And modders are getting surprisingly good at it.

So much so that they found a way to get DLSS 5 working on Radeon GPUs. The performance still needs serious work, but that alone raises an awkward question: how much of what makes a GPU “obsolete” comes down to what its manufacturer is willing to support? When people are paying this much for hardware, I think they deserve an answer.

DLSS 5 On AMD GPUs

Last week, I discussed my biggest concern with DLSS 5. Specifically, what happens if developers begin designing games around the expectation that neural rendering will provide part of the finished design? Because at that point, turning it off could mean losing visual detail that the game was built to have.

This is obviously something that Nvidia would love. Imagine a world where one GPU manufacturer isn’t just faster or provides better upscaling. But playing a game on their card completely changes the way it looks. Think PhysX times 1,000.

Well, one modder may have just crushed Nvidia’s hope.

It’s called DLSS-NR on AMD, and it brings DLSS 5’s neural rendering to Radeon hardware. Specifically, their RX 9000 and 7000 cards. It currently targets DX12 games with FSR support, but so far, it’s only been tested in a few games.

Now, the interesting part is that this isn’t something like DLSS 5 or anything like that. The developer describes a new runtime written for AMD GPUs, using Nvidia’s DLSS 5 library. So this is the real deal.

With that said, there is one major issue: performance.

The YouTuber Ancient Gamplay got a chance to test it out, and in Cyberpunk 2077, the RX 9070 XT went from roughly 85 FPS with DLSS 5 turned off to just 11–12 FPS with it enabled. At that point, you have plenty of time to appreciate the improved lighting. Individually. In every frame.

Luckily, there is some good news here.

The developer now reports roughly 33 FPS at 1080p on the 9070 XT, with further optimization underway and a goal of approaching RTX 5070 Ti performance. That’s obviously just a target, so it’s not guaranteed or anything, but even getting close to that framerate would prove that AMD can easily do the same thing. Even so, the fact that this works at all is huge.

Especially if Nvidia can’t find a way to shut it down. Because let’s be honest, if they can, they almost certainly will. This is clearly meant to be the ultimate walled garden, pretty much forcing gamers to go with Nvidia cards. In fact, this could be bigger than anything Apple, Amazon or any other company has tried in the past. Time will tell how it ultimately ends up, but here’s to hoping competition doesn’t get left in the dust.

RTX 40 & 30 GPUs Just Got Multi Frame Gen

Speaking of software limitations that are clearly meant to force customers into buying the latest cards, it always annoyed me that Nvidia’s 40 series GPUs didn’t get support for Multi Frame Generation. Just like their 30 cards didn’t get support for regular Frame Gen. Obviously there are hardware differences, but AMD has proven that a lot of those can be overcome. Though with some drawbacks.

Well, if this isn’t proof that these limitations are nothing but artificial, I don’t know what is. Because modders have now gotten Nvidia’s own Multi Frame Gen working on RTX 40 cards. And I don’t mean some variant of AMD’s FSR or using Lossless Scaling. This actually uses Nvidia’s official DLSS Multi Frame Generation.

This one is called Universal RTXMFG, and it not only works with Multi Frame Gen, but also Dynamic Frame Gen. At least in certain games. They’ve also added experimental RTX 30 support, though it’s extremely early to the point where some games may load the menu, but still not work. Regardless, this is still really awesome to see.

With all of that said, it looks like Nvidia might be trying to shut them down with some new runtimes that they released. Now, I say “might” because obviously, updates can break unofficial mods without the company specifically targeting them, but this does at least point that possibility. The good news is that they’ve already updated the mod to get around it. And the even better news is that the mod itself seems to work really well.

For example, one test compared an RTX 4080 using the mod against an RTX 5070 Ti with regular Multi Frame Gen support, and the results were scary close. Performance gains and latency were both very similar, and while the 50 series card did have slightly better frame pacing, the reviewer said that they couldn’t actually tell the difference while playing the game.

There are even reports of a 4080 going from roughly 38 FPS with normal Frame Gen to around 180 FPS with 6X Multi Frame Gen. Even though the base rendered framerate dropped to around 30 or 31. So there is some overhead, and Blackwell may still handle frame pacing a bit better thanks to its hardware improvements. But the fact that the RTX 40 cards can run Multi Frame Gen this well makes it a lot harder to believe that support was impossible. Instead, it looks more like Nvidia drew an artificial line in the sand to make their new cards seem more necessary.

Dual GPUs Could Actually Come Back

Believe it or not, dual GPU gaming could be making a comeback! Yeah, I never thought a phrase like that would be coming from my mouth. Or fingers.

If you were around for SLI or CrossFire, you probably remember both the excitement… And subsequent disappointment. Mostly because support for the tech had to be programmed into the game, so it made performance hit or miss. Sometimes a second card gave you a huge FPS improvement. Other times, you spent a bunch of money to discover that your favorite game didn’t particularly care. The tech was also plagued with frequent glitches, weird scaling when going from two to four GPUs, visual artifacts, and just too many annoyances to make it worth doing. Hence why the tech didn’t last.

Luckily, this new approach is completely different. For one, it’s made to offload a specific task onto the new GPU. And second, it’s specifically for DLSS 5 instead of general gaming.

It’s a new mod called Neural Coprocessor, which allows one GPU to run the game while a second GPU handles DLSS 5’s neural processing. Basically, the first card renders the frame, then the second card applies the neural rendering pass and displays the result. That means the main GPU doesn’t have to spend as much of its time doing both jobs.

The developer tested this with two RTX 5060 Ti 16GB cards in The Blood of Dawnwalker at 1080p. He used different upscaling modes, but as an example, in DLSS Quality mode, the game ran at around 98–99 FPS with neural rendering disabled. Enabling it on the same card dropped performance to 54–55 FPS. But moving the neural processing to the second GPU brought the reported game frame rate back up to 91 FPS. That’s a pretty substantial recovery.

Unfortunately, there’s one detail that makes these results more complicated than they initially look. Those numbers measure how fast the game is running. They don’t necessarily show how many finished, neural enhanced frames are actually reaching your monitor.

See, the second GPU still has to keep up. If it can’t process everything the first card sends over, the mod can skip older frames and move on to newer ones. So the game could report 91 FPS without you actually seeing 91 different neural rendered frames per second. Ultimately, we still need measurements of the final output and the full delay to know exactly how much better the experience is.

There are some other issues as well. The developer specifically uses two monitors, with one connected to each card. Running the second GPU without its own display reduced throughput in testing. Plus, this ReShade implementation doesn’t have access to all the engine information that a native DLSS 5 integration would receive.

At the end of the day, this is still very experimental. But it is at least an interesting concept, and if DLSS 5 really does become ubiquitous, it could bring back dual GPUs in a way we haven’t seen in years.

Nvidia’s Melting Connector Problem Is BAD

Unfortunately, it’s not all good news today, as the melting 12V-2X6 connector is catching fire AGAIN! This time, it happened to the prominent reviewer Hardware Unboxed.

In a new video, Steve from Hardware Unboxed noticed that something was wrong after burning his hand on the cable coming from the card, an ASUS ROG Astral RTX 5090. He then checked it with a thermal camera, and the temperature eventually reached over 175 °C. On a power cable.

Using Thermal Grizzly’s WireView Pro, they also found a massive imbalance in how the current was distributed. Some conductors were carrying very little amps, while one reached a whopping 22 amps.

But here’s the interesting part: the worst damage was actually at the power supply end. The cable had melted into the PSU connector badly enough that they needed pliers to remove it. Meanwhile, the connection on the graphics card itself was relatively undamaged. Now, before anyone asks, Hardware Unboxed also said that the connector had been fully inserted.

At this point, all I can do is wonder why Nvidia hasn’t said a word about this. They told us that this issue was fixed with the RTX 50 series, but it’s clear that isn’t the case. Other reviewers have also seen their GPU’s connector melt, so it’s clear that this isn’t just a couple of one offs caused by a connector not seated properly. Something has to be done.

Quick Bytes

Nvidia Finally Gives Us A ReBAR Switch

Nvidia App 11.0.9 adds per-game Resizable BAR controls directly to the app. It also brings DLSS 4.5 Ray Reconstruction support. More settings we can actually control? I’m certainly not complaining.

GPU Shipments Are Up. Desktop Isn’t.

Jon Peddie Research says PC GPU shipments increased 10.4% last quarter to 75.5 million. That sounds pretty good, until you see that notebook graphics shipments grew 16.8%, while desktop graphics actually fell 4%. Those totals also include integrated graphics, so we aren’t talking about 75.5 million standalone gaming cards. Basically, parts of the PC market are doing fine. The people trying to afford a new desktop are having a slightly different experience.

Your SSD Has Its Own Special Cable Now

Lexar announced a portable SSD called the Muse, which it claims is the thinnest in the world. It measures just 3.8mm at its thickest point, with read speeds up to 1,050MB/s and 512GB or 1TB versions planned for Q4. Unfortunately, making it that thin apparently meant using a proprietary magnetic connection. It looks pretty neat, but I already know I’m going to lose the cable and discover that none of the other 400 cables in my drawer will work.

Framework Is Actually Refunding People

Remember when Framework’s LPCAMM2 memory prices went through the roof? Well, the company found a limited supply of cheaper 32GB and 64GB modules and is passing the savings along. Eligible pending orders through Batch 10 will get lower prices, while qualifying orders that already shipped will receive refunds. Prices still aren’t back to their original levels, but giving money back to people who already paid is something I’ll absolutely give them credit for. Plus, they obviously can’t help these absurd prices.

That’s all the bytes for this week.

Until next time, may your GPU get better with age, your RAM come with a refund, and your power cable stay below pizza temperature.

In honor of Meld Bytes success over the last few weeks, I’ve decided to run a BIGGER GIVEAWAY!

Prizes

Three separate winners will be selected. Each winner will receive one of the following prizes:

1. G.Skill Trident Z5 Neo 42GB DDR5-8000 Memory Kit

2. Patriot 16GB DDR4-3200 Memory Kit

3. Samsung 990 PRO 1TB NVMe SSD

Each prize will be awarded to a different winner. Winners may not choose or exchange their prize unless otherwise stated by Gamer Meld.

Eligibility

The giveaway is open to legal residents of the United States who are 18 years of age or older at the time of entry.

Employees, contractors, or immediate family members of Meld Bytes/Gamer Meld are not eligible to enter.

Void where prohibited by law.

How to Enter

To enter, follow the giveaway entry instructions provided by Meld Bytes.

No purchase is required.

Limit one entry per person, unless additional entry methods are specifically stated as part of the giveaway.

Duplicate, fraudulent, automated, or otherwise invalid entries may be removed at the discretion of Meld Bytes.

Giveaway Period

The giveaway begins on 09/05/2026 and ends on 10/05/2026 at 11:59:59 PM Eastern Time.

Entries received after the deadline will not be eligible.

Winner Selection

Three winners will be selected at random from all eligible entries after the giveaway period ends.

Each selected entrant will be assigned one of the three available prizes. Once an entrant has been selected as a winner, they will not be eligible to win another prize in the same giveaway.

The odds of winning depend on the total number of eligible entries received.

Winner Notification

Winners will be contacted using the email address or contact information provided with their entry.

Each winner will have 24 hours from the time the notification is sent to respond and provide any information reasonably required to verify eligibility and ship the prize.

If a selected winner:

• Does not respond within 24 hours;

• Cannot be contacted;

• Is found to be ineligible;

• Provides invalid or incomplete information; or

• Declines the prize;

Meld Bytes may select a replacement winner at random from the remaining eligible entries.

Shipping

Standard shipping of the prize to an eligible United States address will be provided at no cost to the winner.

Meld Bytes is not responsible for incorrect shipping information supplied by a winner, delivery delays, lost packages after confirmed delivery, or other circumstances outside its reasonable control.

Prize Conditions

Prizes are awarded as-is and may not be redeemed for cash.

No substitution or transfer of a prize is permitted except at the discretion of Meld Bytes. If a listed prize becomes unavailable, Meld Bytes reserves the right to substitute a prize of equal or greater approximate value.

Any taxes or other obligations resulting from receipt of a prize are the responsibility of the winner where required by law.

General Conditions

By entering, participants agree to abide by these rules and the decisions of Meld Bytes regarding the administration of the giveaway.

Meld Bytes reserves the right to disqualify entries involving fraud, abuse, manipulation, automated entry methods, or attempts to circumvent these rules.

Meld Bytes also reserves the right to modify, suspend, or cancel the giveaway if circumstances outside its reasonable control prevent it from being conducted fairly or as intended.

Platform Disclaimer

This giveaway is sponsored and administered by Meld Bytes/Gamer Meld unless otherwise specifically stated.

It is not sponsored, endorsed, administered by, or associated with Samsung, G.Skill, Patriot Memory, YouTube, Substack, or any other third-party platform or manufacturer unless expressly stated otherwise.

By entering, participants acknowledge that any information submitted in connection with the giveaway is being provided to Meld Bytes and not to those third parties.

Don’t worry though, because Nvidia just released their new DLSS 5. And while it honestly looks pretty good in most cases, the performance hit is pretty drastic. But that isn’t even the biggest issue.

Luckily, I actually have some good news about an oversupply, there’s new competition and AMD is going all out!

DLSS 5 Changes What A Game Even Is

Nvidia’s new technology, called DLSS 5, is officially here. And with it, brings the potential to completely change the entire industry. But the problem with talking about “change” is that most people tend to forget that change can be bad. In this case, it could redefine what a game even is. And in the end, we may not have a choice.

So what exactly is DLSS 5? Well, Nvidia calls it 3D Guided Neural Rendering. Unlike previous version of DLSS, which all focused on raising performance, DLSS 5 actually changes the final image itself. Nvidia research describes it as “the first DLSS technology to generate the final displayed appearance rather than reconstruct a higher-cost reference output from the conventional renderer.”

Basically, DLSS 5 takes information from the engine like color, motion vectors, materials, surface information and lighting and feeds it into an AI model. That model can add things like more realistic skin, light passing through the ears and hair, deeper contact shadows, improved material responses and more complex lighting. And to be fair to Nvidia, this isn’t like typing a prompt into an image generator and hoping your character still has the same number of fingers in the next frame. The system is designed to be deterministic and stable.

Unfortunately, this brings us to our first problem. While Nvidia claims that developers have control over what the AI outputs, they really don’t. They can choose between different neural models, adjust the Intensity, apply semantic masks, and isolate individual assets or groups of objects so DLSS only modifies specific parts of the scene. But they can’t make any specific, meaningful changes to the final output. What I mean is, they can’t talk to it like ChatGPT. So if the AI decides to change something in a way that the developer doesn’t want, they can’t just tell the AI to change it to something else. We’ve already seen this issue play out with character models that end up looking completely different from their original design.

Sure, you could just lower the intensity, but then you end up with a mismatch of some objects that have great lighting and all that DLSS 5 offers and some without.

This also brings up the issue of every game looking the same. As if we don’t already have a major problem with Unreal Engine games that tend to have a very similar style. Now, you’re adding another layer on top of that. Not to mention the argument against games becoming so real that they’re not an escape anymore. That one, I don’t fully agree with, but it is nice to see games that have their own style or design. Not everything has to look like real life.

It EATS GPUs

When it comes to actual performance, Nvidia’s DLSS 5 destroys your GPU. Luckily, it’s not as bad as it was when they first demoed the new tech back in March. Remember they were using 2 RTX 5090s for that. According to Nvidia, it’s now 5X faster than that demo, but unfortunately, there’s still a massive performance penalty.

According to Nvidia themselves, you can expect between a 50-60% performance hit when turning it on. But when they actually show the performance in NBA 2K27 (the first game to get DLSS 5), Nvidia makes it look way better than what it really is.

As you can see, the RTX 5090 gets a very impressive 370 FPS in 4K. And the 5080 still gets a respectable 233. BUT if you look down at the bottom of the image, you’ll likely notice something important. This is using 6X Multi Frame Gen! So not Dynamic Frame Gen, where it dynamically changes the number of AI-generated frames between each real frame as they’re needed. Nope. This gets the full 6X. When we do the math, you’ll quickly realize that the 5090 is getting just 62 real frames per second, while the 5080 gets a measly 39. Not to mention that it’s also using upscaling set to Performance mode.

Now, the upscaling part really isn’t the worst of it, because at least with upscaling, you can actually get lower input latency by increasing the number of real frames being rendered. But with frame gen, you don’t get that same benefit. Basically, those AI-generated frames can’t incorporate new input from your mouse or keyboard. The game has to wait for the next real frame to actually respond to what you’re doing. So while the game might look like it’s running at 233 or even 370 FPS, it absolutely doesn’t feel like it. But like I said, the terrible performance isn’t the real problem.

The Real Problem

The biggest reason this isn’t necessarily a problem is the obvious one many of you will likely point out: DLSS 5 is completely optional. If you like the way it looks, turn it on. If you don’t, leave it off. And to be fair, some of the results really are impressive.

BUT while this answers the issue of performance, it also points to the real problem: Super Resolution was also optional.

It originally gave developers a way to take a game that already ran well and make it run even better. But over time, that slowly shifted. To the point where games seem to be designed around the expectation that upscaling would already be enabled.

Instead of optimizing a game to hit 70 FPS natively and letting DLSS push that to 100+, suddenly 70 FPS is only possible with DLSS enabled. So it stopped being the bonus and became part of the performance budget.

And DLSS 5 could easily run into the same problem.

If a developer knows DLSS 5 can create convincing skin shading, ambient occlusion, lighting and material detail afterward, why spend as much time and money building all of that through traditional rendering?

Obviously, Nvidia says DLSS 5 is meant to extend the rendering pipeline rather than replace it, but the economic incentives are pretty apparent. And if the base version of a game eventually gets designed with the assumption that neural rendering will provide a bunch of the final visual detail...

What exactly does the game look like when you turn it off?

An Oversupply Could Finally Be Coming

Believe it or not, I finally have some genuinely good news. Not “prices increased slightly slower this week” good. But actual good news.

Because according to TrendForce, the NAND Flash shortage may finally have an expiration date. For anyone who doesn’t know, NAND is the type of flash memory used inside SSDs. And like basically everything else that AI servers need right now, NAND has been getting hammered. TrendForce estimates that the NAND market will end 2026 with roughly a 4 to 5 percent supply deficit. So unfortunately, we still have to get through the current mess.

But in the second half of 2027, TrendForce expects supply to exceed demand. In other words, we could go from not having enough NAND... To having more than the market needs. And there are multiple reasons why.

First, manufacturers are finally beginning to expand again. An August TrendForce report claims that major NAND suppliers are ending their aggressive capital expenditure cutbacks and restarting longer term expansion plans. That means more advanced manufacturing processes, higher capacity products and even entirely new fab capacity coming online.

At the same time, consumer demand is terrible.

TrendForce expects smartphone production to fall somewhere around 15 to 20% this year, while notebook shipments are expected to drop by roughly 10%. And despite all the attention on AI servers, smartphones and notebooks still account for nearly 40% of total NAND demand.

Then there’s China.

The Chinese NAND manufacturer, YMTC is apparently planning to go all out. Because they’ve reportedly set their sights on becoming the world’s largest NAND supplier by the end of 2027. What’s wild is that to accomplish this, they have to nearly double their market share in just 16 months time.

So when you put all of this together: more fabs, more output per wafer, more Chinese capacity, and weak consumer demand, you get a recipe for turning a shortage into oversupply. And that’s when prices really get good.

China Just Beat The Memory Giants To LPDDR6

And speaking of more competition from China, CXMT just pulled off something that I don’t think many people expected. The Chinese DRAM maker has just become the first manufacturer to publicly confirm both LPDDR6 volume production and an actual device that’s using it. Specifically, it’s set to appear inside Xiaomi’s upcoming 18 Fold smartphone.

Now, I will say that SK Hynix has already developed and validated LPDDR6, and they previously said that they planned to begin supplying the memory during the second half of this year. So it’s not like CXMT invented LPDDR6 while Samsung and SK Hynix were asleep. But what they did do, was reach actual commercial production before the biggest companies out there.

Not only that, but their new LPDDR6 parts reportedly reach up to 12,800 Mb/s and capacities of up to 16GB. LPDDR6 itself is a pretty big upgrade over LPDDR5 memory as well, moving to PAM3 signaling and providing much more bandwidth. But the main story here is really the speed at which CXMT is catching up.

It was just a few weeks ago that we were talking about CXMT DDR5 getting support for higher frequencies. Then it started appearing in memory products and laptops. Then reports showed their overall production getting much closer to the more established manufacturers. And now it’s at the forefront of a brand new memory generation.

Of course, there are still some big, unanswered questions. Getting LPDDR6 into one flagship phone does not prove that CXMT can manufacture it at Samsung or SK Hynix scale. We don’t know how its yields will compare, how much capacity it has available or how quickly it can expand production. At the end of the day though, it’s great to see new competition.

AMD’s Going After Nvidia’s Biggest Asset

There’s also a completely separate story about competition that’s happening right now. And this one could have huge ramifications for the entire industry. Because it may finally mark the end of Nvidia’s monopoly. AMD just announced ROCm 10.

For those who don’t know, Nvidia’s biggest advantage in the AI market isn’t actually their hardware. But their software. Specifically, CUDA. See, Nvidia has spent nearly two decades building a massive software ecosystem around their hardware. Frameworks are optimized for it. Developers already know it. Companies have built entire infrastructure stacks around it. So even if AMD builds an accelerator with competitive performance (which they mostly have already), convincing a company to rewrite and revalidate their software stack is an entirely different problem.

That’s where ROCm 10 comes in. One of the biggest additions in ROCm 10 is ROCm.AI, which includes a system called Hyperloom. Hyperloom can actually look at an AI workload, find bottlenecks, modify things like kernels, memory management and scheduling, benchmark the changes and then validate that they actually work. Basically, AMD is using AI to optimize software for AMD’s AI hardware.

And AMD is showing some huge performance gains. For example, on the same eight GPU MI355X platform, a preview of ROCm.AI delivered an average 3.3X higher inference throughput across three tested models and 2.4X higher training throughput compared with ROCm 7.

With that said, AMD explicitly warns that these workloads are not universal ROCm 10 performance jumps. But the improvement shows what AMD believes it can gain by attacking the entire software stack instead of relying on faster silicon alone.

Not only that, but ROCm 10 also brings AMD-specific workflows directly into coding tools like Codex, Claude Code and Cursor, while expanding support across major AI frameworks and tools.

Basically, this is exactly what AMD needs to do if they ever want to really compete with Nvidia. They need competitive hardware that developers don’t hate using.

Quick Bytes

920Hz Lasted Four Days

In the last issue, I told you that ASUS had announced a monitor capable of hitting 920Hz. Well, congratulations to anyone who bought it. Your technology is now ancient. Acer just announced the Predator XB253Q U1, a 24.5-inch 1080p IPS gaming monitor capable of running at 1000Hz.

One. Thousand. Hertz.

At this rate I’m expecting 1440Hz by Tuesday. And just like I said last week: if you still miss the shot, I really don’t know what else the industry can do for you.

Microsoft Is Turning On Another Security Feature

Starting in October, Microsoft will begin automatically enabling Windows 11’s Memory Integrity protection on a broader range of eligible PCs. Memory Integrity uses virtualization based security to isolate important parts of the Windows kernel, which is obviously good for security.

Unfortunately, Microsoft has previously acknowledged that it can reduce gaming performance on some systems, with older hardware generally taking the larger hit.

The RTX 5090 Is A Meme Now

Nvidia’s RTX 5090 has actually reached $5,090. And it’s not some liquid-cooled collector card either. PCPartPicker has actually listed the regular ASUS TUF RTX 5090 OC at $5,099, which is now more expensive than MSI’s ridiculous $5,090.99 Lightning Z. The jokes write themselves.

Nvidia’s Newest Driver Has A Problem

GeForce 616.56 has caused black/white flashes, brightness changes and color corruption for some GTX 16 through RTX 50 users. Nvidia says it’s testing a fix; switching from 8-bit to 10-bit color reportedly fixes it for some people.

That’s all the bytes for this week.

Until next time, may your frames be real, your RAM be real, and your GPU cost less than your entire PC.

Of course, we’ve had multiple cycles over just the last few years, but the AI boom may just outlive them all. Luckily, AMD’s next generation CPUs and GPUs are starting to look absolutely ridiculous, so there’s at least something to look forward to.

The Desktop PC Market Is Getting Crushed

Before we get into anything positive though, I have to discuss the ugly. And let’s just say, your retina’s will burn.

According to Mercury Research, desktop CPU shipments fell by more than 20% in a single quarter. What’s crazy though, is that the overall CPU market actually grew by more than 10% during that same period. All thanks to laptops and, I’ll bet you can’t guess… Servers.

Okay, you guessed it. Basically, processors aren’t collapsing everywhere.

But on desktop, they’re getting obliterated.

As for why, I’m sure you’ll guess that as well. Limited GPU availability combined with ridiculously expensive memory and SSDs are making a full desktop upgrade harder and harder to justify.

And this is even worse than what AMD warned about just a few months ago. Back then, Lisa Su predicted a 20% gaming revenue decline in the second half. Instead, prices fell over 20% in just one quarter.

Unfortunately, we’re starting to see the consequences go beyond a bad sales chart.

Australian custom PC builder, Dream PC, recently announced that it was shutting down after almost a decade in business. The company specifically pointed to years of component pricing and availability chaos, from COVID, to crypto, and now the enormous demand created by AI.

Don’t worry though, because it doesn’t stop there. As Pine64 also announced that they’ve stopped production across their entire Linux hardware lineup because of DRAM and eMMC shortages. The company currently doesn’t expect production to restart until sometime after the middle of 2027.

And unless memory pricing turns around faster than everyone’s saying, I highly doubt they’ll be the last.

2027 Is The Year of AMD

Thankfully, the future of desktop hardware looks considerably more interesting than the present. Because both AMD’s next generation Ryzen CPUs and Radeon GPUs are gearing up to be their biggest upgrades EVER!

Ryzen 10,000 (PLEASE don’t call it AI)

The upcoming Zen 6 desktop family, codenamed Olympic Ridge, are confirmed to be built on TSMC’s 2nm process. For those who didn’t know, that’s two node jump from Zen 5’s 4nm process.

But that’s just the beginning, as one leak claims that AMD is planning to push the clocks all the way up to 7GHz! In fact, a source from Moore’s Law Is Dead claims to have documentation that references eventual qualification samples actually hitting that clock speed. Meaning AMD really expects these to do it. For comparison, that’s an unreal 23% faster than Zen 5’s 5.7GHz. If this ends up being true, it’ll be the biggest clock upgrade since 1st Gen Ryzen.

Not only that, but according to another leak, we could be looking at a 10%+ average IPC increase over Zen 5. Don’t forget that IPC (Instructions Per Clock) doesn’t account for any clock increases, core count increases, etc. It’s just how much faster the CPU is at the same clock. And it’s very application specific because it’s ultimately dependent on the architectural changes from one architecture to the next.

And the upgrades keep on coming, as multiple leakers claim that Zen 6 is adding more cores! Specifically, it’s set to get as many as 24 cores, versus 16. Making for an increase in core count by as much as 50%. And it’s all thanks to the chiplets that make up the Zen architecture going from 8 to 12. Meaning the entire product stack is going up in cores. Which is huge news for AMD’s X3D chips. Because it means that the 3D V-cache chiplets would come with 12 cores instead of 8. Think the Ryzen 7 10800X3D.

To top it off, all of this is officially coming to AM5 motherboards. AMD themselves recently listed Olympic Ridge with AM5 support. Not only that, but their APUs, codenamed Medusa Point, also appeared with both mobile FP10 and AM5 support. So AMD’s clearly planning to bring their Zen 6 based APUs to desktop. Of course, none of this guarantees that every existing AM5 motherboard will get Zen 6 support, but the platform itself is sticking around for yet another generation.

RX 10,000

Moving over to GPUs, AMD’s next generation Radeon cards are set to be based on either their RDNA 5 architecture, or possibly UDNA. Remember that AMD confirmed a little while back that the company was moving over to a unified architecture, with both their gaming RDNA and compute CDNA becoming one UDNA. And according to one leak, RDNA 5 is already based on UDNA, but with some tweaks made for the PS6. It sounds like it could still be called RDNA 5, but according to the leak, it’s still ultimately based on UDNA.

And that architecture helps to explain the incredible 154 CUs claimed to be in the top end card. That’s a whopping 2.4X the cores than AMD’s current Gen RX 9070 XT. And the well known leaker, Kepler, actually claimed that the top end AT0 die will 100% release for consumers. As long as it does actually release.

Finally, AMD themselves have confirmed some incredible stuff with Project Amethyst. They claim themselves that it’s set to go in future hardware, so we can definitely expect it in RDNA 5. Like their new dedicated hardware block, called a radiance core. This is made specifically for ray traversal, so we can expect some big performance jumps out of AMD’s next Gen ray tracing. They also announced Neural Arrays and Universal Compression. Ultimately, 2027 is set to be the year of AMD. And both Intel and Nvidia need to watch out.

Nvidia’s RTX 5000 Performance Is Really Weird

Nvidia’s RTX 50 series has a pretty major problem. And it can be easy to miss if you only look at traditional GPU reviews.

In certain games, especially as you move down from native 4K toward 1440p or 1080p, the RTX 50 series can lose a shocking amount of its advantage. There are even individual examples where AMD’s RX 9070 XT can match or beat an RTX 5080.

For example, In Call of Duty: Black Ops 7 at 1440p, the 9070 XT beats the 5080. In Borderlands 4 at 1080p, the 9070 XT and 7900 XTX can effectively match the 5080. And Space Marine 2 produces one of the strangest examples of all. At 4K, the RTX 5090 is significantly faster than the RTX 4090. Move down to 1080p, though, and the 4090 actually crushes the 5090.

Now, some amount of shrinking GPU performance differences at lower resolutions is completely normal. The lower the resolution, the more your CPU becomes the bottleneck. But some of these results clearly go well beyond the normal scaling you would expect.

As for why, there are a few theories. With the biggest being Nvidia’s shift in hardware scheduling by adding the new AMP or AI Management Processor. AMP basically adds a layer between the CPU and gigathread engine, which some people think might be adding latency if the workloads aren’t optimized for it. Another is Nvidia’s greater driver overhead in CPU limited scenarios. Ultimately, there doesn’t appear to be one single proven explanation covering every game.

And believe it or not, this actually matters even if you bought a 5090 specifically for 4K. Because a lot of people use upscaling.

Specifically, even DLSS Quality at 4K internally renders at 1440p before reconstructing the final image. If you move down to balanced or performance, the internally rendered resolution gets even worse. So if a GPU behaves strangely as internal resolution drops, native 4K benchmarks may not tell you exactly how the card will behave when you actually use DLSS.

Basically, before dropping thousands of dollars on a GPU, look at benchmarks using the settings and resolution you actually plan to play at.

Microsoft Really Wants RTX Spark To Work

At the same time, Microsoft is putting a very unusual amount of effort into Nvidia’s upcoming RTX Spark PCs.

In a new, experimental Windows 11 build, it includes a hidden feature called IntelligentCarveout. This would essentially allow systems with unified memory to reserve part of that memory specifically for graphics and AI workloads. That’s important because RTX Spark combines its CPU and Blackwell GPU with a shared memory pool that can reach 128GB. Basically, instead of the CPU having one pool and the GPU having its own, everything shares the same memory.

What’s wild is that Microsoft already changed Windows earlier this year to allow RTX Spark’s GPU to access a much larger portion of that unified memory. Now, it’s apparently working on actual user facing controls around the same architecture. And it’s not so much the changes that Microsoft made that interests me. Rather, it’s the fact that Microsoft made them in the first place.

Let’s just say that the OS maker doesn't usually make multiple fundamental memory management changes for some random niche PC nobody expects to sell. So they clearly expect Nvidia’s new platform to matter.

Whether that means RTX Spark is incredibly good or Nvidia is simply incredibly persuasive remains to be seen.

Probably a little of both.

Quick Bytes

CPU-Z Just Got Its Biggest Upgrade In Years

CPU-Z 3.01 is out, and CPUID has massively expanded what the program can actually do. Its new Validator V3 can run more than 100 system health checks in under 10 seconds, looking for things like outdated drivers, overheating hardware and suspicious system configurations.

There’s also a new validation system for extreme overclocking, and a more advanced mode is being developed with stress testing and deeper CPU, GPU, RAM and storage diagnostics.

The RTX 3060 Somehow Got MORE Expensive

Remember Nvidia brought the 12GB RTX 3060 back because the entry level GPU market needed more options? About that…

An MSI RTX 3060 12GB that relaunched at Newegg for around $330 in July is now listed at $480. That’s a 45% increase in less than two months. That’s actually more expensive than when it first released back in 2021!

Intel Built A CPU Around Being Cheaper

Intel just revealed more about their upcoming Wildcat Lake. The architecture behind its mainstream Core Series 3 laptop chips, designed specifically for cutting costs.

Compared to Panther Lake, Intel cut the compute die by 38%, reduced the P-core count, halved the GPU configuration, shrank the NPU and replaced its more expensive Foveros packaging with an organic UCIe design.

Intel says Wildcat Lake is meant for price-sensitive laptops, with systems starting around the $450 range.

920Hz Monitors Are Apparently Necessary Now

ASUS recently announced a new gaming monitor capable of running at up to 920Hz. You heard that right. Nine hundred and twenty!

The 27-inch ROG Strix XG27ACNS uses a Fast IPS panel and can switch between QHD at 475Hz, 1080p at 620Hz and 720p at 920Hz. ASUS also announced a 720Hz OLED esports display at Gamescom.

Now, I’m not saying 920Hz is unnecessary. I’m just saying if I still miss the shot, we are officially out of hardware to blame.

That’s all the bytes for this week.

Until next time, may your socket survive another generation, and your five-year-old GPU never relaunch for more than it cost new.

PS: I apologize for being so late this week. Between getting COVID and someone trying to steal my catalytic converter, it’s been a tough one. I definitely hope to be on time from now on.

Unfortunately, the answer looks to be: A LOT!

New reports are warning us that DDR4 in particular is about to get way more expensive. And even parts of the market that remained relatively “cheap” are now getting dragged upward. The AI boom is starting to completely distort the economics of the entire PC industry.

Luckily, there’s a new technology that gives us real hope. Because it isn’t about shifting memory production around from one type to another, or producing more of the same. Instead, it’s actually made to replace it entirely! At least enough of it to actually help ease the burden of outrageous demand.

RAM Prices Have Completely Lost The Plot

Before I get into the actual solution, we have to talk more about the problem. Because there absolutely is one, and it’s getting much worse.

In a new report, Tom’s Hardware went through data from PCPartPicker, comparing prices between August 2025 and August 2026. A 64GB DDR5 5600 kit that averaged around $191 last August is now roughly $1,118. DDR5 6000 went from around $222 to $1,272. That’s an increase of roughly 485% and 473% in one year.

And it gets even worse! When we compare the lowest cost the kits have ever been, one 128GB DDR5 6400 kit currently sits at an unreal $3,399. Yet the lowest price Tom’s Hardware ever tracked for that capacity and speed was $329.

That’s more than 10X the cost!

Even DDR4. The thing everyone was supposed to fall back on, is getting hammered. Depending on capacity and speed, Tom’s Hardware found an increase of between 120% and 177% year over year. Don’t worry though. It only gets much worse!

A recent Morgan Stanley report expects DDR4 prices to jump another 50% in Q3, followed by more than 10% in Q4. As for why, Morgan Stanley claims that manufacturers are moving capacity over to HBM as well as newer DRAM. Luckily, one of those is DDR5, but I highly doubt it will get much of the capacity. Instead, most will likely transition into HBM. And obviously that makes perfect financial sense, but it sucks for basically everyone else.

Share

There’s An Actual Solution!

This is where things get interesting. Sandisk and SK hynix just announced a brand new memory technology called HBF, or High Bandwidth Flash. The “Flash” part is the key. And it’s exactly what you think! High Bandwidth Flash is essentially made from NAND Flash. The same technology that’s used inside your SSD. And it’s being turned into an entirely new class of memory!

This means that companies can use NAND flash as memory in AI workloads. That could obviously raise prices of SSDs, but because NAND flash is much cheaper and easier to scale into huge densities, it should finally meet a lot more of the demand. Which obviously means lower prices for everyone else.

Now, if you know anything about the technology, you’re probably asking exactly what I did when I first heard about it: how?

For one, HBM4 can reach bandwidths of up to 4.1 TB/s. Compare that to a PCIe Gen 5 SSD, which peaks at around 14-15GB/s. Not to mention endurance and latency. So there’s obviously a huge disparity here. And that’s where this gets interesting. To get around this, they essentially took the same approach as HBM, but applied it to NAND flash. HBF actually stacks a huge number of NAND dies together, and can access them all in parallel. It also uses the industry standard UCIE interconnect. A fast interconnect for die to die communication. This gives HBF bandwidth at up to 3 TB/s.

Sandisk has previously said that HBF could eventually offer roughly 8 to 16 times the capacity of HBM at a similar package cost. Basically, this could make a massive difference when it comes to memory supply.

With that said, it does still suffer from high latency, as well as lower write endurance. To manage that, they use HBF for strictly read dominant AI inference workloads. so there’s still a need for at least some HBM for certain tasks. But, given inference is exploding past training, it’s becoming the more dominant workload anyway.

The biggest issue, is really the timing. Because HBF has to be made into the AI accelerator, similar to HBM, companies have to build them around the new memory. The good news is that Sandisk actually claims that HBF equipped AI hardware is targeting early 2027. That may sound a long way off, but remember that companies expect this crisis to take years to end. So 2027 isn’t all that bad. Either way, this look to be a fantastic idea for actually solving a problem that looked unsolvable.

Now Budget GPUs Are In Trouble Too

Unfortunately, there is still bad news to cover. Specifically, in regards to graphics cards.

Because PC Partner, the company behind brands like Zotac and Inno3D, is warning that entry-level GPUs could face even more severe shortages during the second half of 2026. And they specifically warn that those shortages could push average selling prices even higher.

That’s especially concerning because cheap GPUs were basically the last somewhat-normal part of the market. PC Partner’s own brand GPU sales volume has already fallen by 18.4% year over year, while its average selling price actually increased by 10.7%. Meaning they’re selling much fewer cards, for more money. And it doesn’t stop there. According to PC Partner, there are longer lead times for CPUs, system memory, and other components.

And this actually starts to make sense out of AMD’s new RX 9050 release. As many of you know, the company launched their 8GB card for around $279. Yet the card is terrible at that price! The RX 7600 was more than 20% faster at 1080p and even the RTX 5050 makes this look terrible.

At today’s prices, I don’t think the 9050 makes much sense. But I increasingly wonder whether AMD priced it for the GPU market it expects six months from now instead of the one we have today. Which isn’t exactly comforting, given prices are already terrible.

Intel Actually Wants Desktop Users Again

While the component market catches fire, CPU competition is getting a lot more interesting. Intel VP, Robert Hallock, was recently interviewed by Tom’s Hardware. And he made something pretty clear: Intel wants Nova Lake to make a statement on desktop.

As many of you likely already know, AMD is launching its new Zen 6 architecture in EPYC servers first. That’s pretty odd for them because they typically release to desktop first. And Mr. Hallock was asked about this.

According to him, Intel is debuting their new core on desktop first. More importantly, Intel says it has reorganized its enthusiast desktop business with dedicated people, leadership, product managers and budget, while planning its desktop roadmap all the way through 2030.

But the really big news, is that Mr. Hallock described Nova Lake as potentially one of Intel’s most significant desktop CPU launches ever.

Now, corporate promises don’t mean much, but it does sound like Intel is going all out for next Gen. Arrow Lake didn’t exactly leave Intel with an unlimited supply of goodwill, and Nova Lake still needs to physically exist and beat AMD before anyone starts handing out trophies.

But this is exactly what I want to hear. Intel got punched in the face. Hard. But instead of abandoning desktop PCs for AI accelerators and servers, it’s apparently putting more resources into trying to win enthusiasts back.

Competition works.

Who knew?

Meanwhile AMD Is Already Preparing Zen 7

With that said, AMD isn’t just waiting around for Intel to punch. The company is already working on their next generation Zen 7 architecture. Well, they’re technically always working a couple of generations ahead, given how long a new architecture takes. BUT they’re really making progress.

Because the newest AIDA64 beta already lists preliminary support for AMD Zen 7 server processors.

That’s right. Zen 7. We’re still talking about the Zen 6 launch, and AMD’s already long into development of their follow-up. AMD has officially confirmed multiple Zen 7 EPYC families for 2028, although the crazy specifications you’ve probably seen floating around are still rumors.

One leak claims a Zen 7 server design could feature an IPC uplift of around 15–24%, TSMC’s A14 process and chiplets containing up to 48 CPU cores each. And the highest end part comes with up to 8 chiplets for an unreal total of 384 cores! Again, that hasn’t been confirmed, but if it does end up being true, Intel really will have to bring the heat. I mean that’s 50% more cores than the upcoming Zen 6 architecture, which already has more cores than Zen 5 Epyc.

So while Intel may finally be gearing up for a serious comeback, AMD is determined to make sure that comeback is as difficult as humanly possible.

Gotta love competition!

QUICK BYTES

NEVER Trust Cloud Providers

Nine PBS recently lost access to roughly 50TB of archival material spanning 70 years after its contracted cloud storage provider, called Open Source Storage, effectively disappeared. What’s wild is that they were actually using another cloud company, called Iron Mountain to store that data. At first, Iron Mountain agreed to give back the data, but they then backed out, arguing that OSS is technically their customer, and they own the data.

There is an update since I last discussed this in a video though. A judge has cleared the station to recover the data from Iron Mountain’s servers under specific conditions. So this may actually have a happy ending. Still, I think the lesson remains the same: The cloud is somebody else’s computer. Keep another copy.

Noctua Still Refuses RGB

Noctua says it still has no plans to add RGB lighting or screens to its coolers and fans. The company would rather spend those engineering resources improving cooling and noise. And I couldn’t be prouder.

Windows Update Is Breaking Games Again

Microsoft is investigating reports that its latest Windows 11 update can cause some games to freeze, crash or even reboot the PC. Apparently Windows Update decided PC gaming wasn't challenging enough already.

Thieves Are Downgrading Gaming PCs

Two thieves in China reportedly stole RTX 5080 and 5070 GPUs from esports hotel PCs and replaced them with cheaper cards like RTX 3060s so nobody would immediately notice. They were eventually caught after a power outage interrupted one of the swaps.

That’s all the bytes for this week.

Until next time, may your RTX 5080 stay an RTX 5080, your 12V-2x6 stay boring, and your cloud provider continue to exist.

Instead, basically every story seems to be yet another sign of just how weird the hardware market has become. Memory prices have gotten so ridiculous that companies are redesigning AI hardware to use less of it. At the same time, Intel is apparently getting back into the memory business. Then a four year old, $80 Ryzen CPU briefly became Amazon’s best-selling CPU, and a ton more.

So yeah. Everything is “normal.”

The RAM Crisis Is Eating Itself

While the memory crisis obviously isn’t over, some of the forces that could eventually mark it’s end are getting much stronger!

For example, TrendForce recently reported that DRAM spot price increases were losing momentum. Mainstream DDR4 spot pricing increased by just 0.17% between July 29 and August 4. Now, I know that’s still an increase, but compared to a whopping 3.27% jump only a couple of weeks earlier, that’s a noticeable slowdown.

What’s even better, is that buying activity has remained incredibly weak over the days since. For example, on August 6, most inquiries were basically just buyers checking prices. On August 7, DDR5 inquiries weren’t really translating into purchases, and on August 10, inquiries were actually slowing down.

Basically, buyers are starting to reach the point where they simply refuse to keep chasing prices higher.

What’s wild is that even AI companies are feeling this way as well. And this includes NVIDIA!

According to another report from TrendForce, Nvidia has begun evaluating lower memory configurations for its future Rubin Ultra hardware, including versions that use fewer HBM4E layers, stepping back to HBM4, or even looking for alternatives. TrendForce also says that cloud providers are considering similar reductions. All while Nvidia already cut the planned SOCAMM capacity of it’s Vera Rubin Superchip modules because of LPDDR5X supply concerns.

Meaning the companies that helped to create the memory shortage, are now looking for ways to use less memory because it’s gotten so expensive.

That’s a big deal.

At the exact same time, manufacturers are pouring money at supply. The Chinese memory maker, CXMT, is now considering to build another 12-inch DRAM fab in Beijing. And this is on top of a huge expansion that’s already underway. They’re expanding so fast, that their new facilities being build could eventually more than double their production from 300,000 wafers per month to 600,000.

But that’s just the beginning. SK Hynix just announced that they’re investing an insane 54 TRILLION into building new memory fabs. Okay, that’s 54 trillion Korean Won, but that’s still roughly 38.3 billion US dollars.

And now things are getting even stranger.

Intel’s CEO, Lip-Bu Tan just revealed that new memory architectures are one of his “pet projects.” He used to be against the memory business because he considered it a commodity, but not anymore. To give you an idea of how serious he is, Intel recently hired the former SK Hynix CEO, Seok-Hee Lee. What’s wild is that Intel actually started out as a memory company before leaving it decades ago. So we’re now going full circle.

Now, this doesn’t mean that memory prices will get cheaper next Wednesday, but with buyers refusing to pay higher prices, AI companies designing products to use less memory, and major memory makers ramping aggressively, things are ripe for a correction.

And it could be cataclysmic!

An $80 Ryzen CPU Proves Just How Bad It’s Gotten

Here’s another great example of what the memory crisis is doing to the PC market.

AMD recently shared a screenshot showing the four year old Ryzen 5 5500 sitting at the top of Amazon’s best selling CPU list, even beating out AMD’s popular X3D processors.

The 5500 has since lost the top spot again, and Amazon’s ranking obviously aren’t the same thing as audited sales data, but the interesting part is why a roughly $80 Zen 3 processor is still anywhere near the top in 2026.

It runs on AM4. And that means DDR4.

Tom’s Hardware noted that, among Amazon’s top 25 CPU listings at the time, nine were Zen 3 processors using DDR4 platforms, compared with 12 Zen 4 or Zen 5 chips using DDR5. That’s kind of incredible when you remember that AM5 launched nearly four years ago.

For years, the normal enthusiast advice was pretty simple: if you’re building a new machine, move to the latest platform.

But when DDR5 costs enough to make you question every decision that brought you to the checkout page, suddenly an $80 processor and some cheap DDR4 you may already own starts to look pretty good.

It means price is becoming a much bigger part of the performance equation.

Your CPU Has A Vulnerability

AMD recently announced two new vulnerabilities that affect just about everyone, including Intel.

Specifically, there are two separate flaws in code that was published by the Trusted Computing Group. And they both involve TPM 2.0. Yes, the same TPM 2.0 that Microsoft requires for Windows 11.

One vulnerability has the potential to expose sensitive TPM information and allow falsified keys, while the other involves a timing side channel that can expose TPM protected data.

AMD confirmed that this vulnerability affects a wide range of Ryzen processors. I’m talking Ryzen 3000, 4000, 5000, 6000, 7000, 8000 and even Ryzen 9000 CPUs. Plus tons of mobile processors, multiple Threadripper generations and even some Epyc chips.

Intel confirmed that a ton of their platforms were affected as well, stretching across numerous Core generations and chipsets.

With that said, there is some good news. The mitigations already exist.

For most desktop users, you want to check your motherboard manufacturer’s support page and see whether there’s a newer BIOS available. If so, you likely want to update it. You can cross reference the firmware with the updates from AMD and Intel to make sure that it includes the mitigation.

Nvidia’s First PC CPU Looks Surprisingly Good

Even though we’ve known about it for quite a while, it’s still odd to think about Nvidia becoming a direct CPU competitor to AMD and Intel. But here we are.

And today, two versions of Nvidia’s upcoming N1X recently appeared in Geekbench 7.

The full model comes with 20 CPU cores alongside the huge Blackwell integrated GPU with 6,144 CUDA cores. Then there’s a cut down version with 18 CPU cores and 5,120 CUDA cores.

The cut down version scored 2,541 in single core and 21,776 in multi core. That’s not bad at all. But obviously the more interesting chip, is their flagship model.

That 20 core chip scored 2,570 in single core and 23,126 in multi core. When comparing this to AMD’s Ryzen AI Max+ 395, Nvidia’s new chip is roughly 6% faster in single core and 12% faster in multi core. Against Intel’s newest Panther Lake Core Ultra X9 388H, Nvidia was roughly 7% behind in single core, but nearly 20% ahead in multi core.

With all of that said, there is one problem with this. It’s Geekbench. So it isn’t the best representation of real world performance, but it can at least be an indication of what to expect. And if this is right, Nvidia’s upcoming CPU is a seriously impressive chip. Time will tell just how far Nvidia can take this.

Microsoft Has Lost It

Remember when Microsoft recommended 32GB of RAM as a “no worries” Windows 11 upgrade? Then they quietly removed it? Well, they’re now pulling even more articles that recommend 32GB of memory for gaming. That’s almost certainly thanks to the fact that their now selling a Surface PC with only 8GB of RAM.

So you’d think Microsoft would seriously start worrying about RAM usage in Windows, right? RIGHT?!

Wrong. Because new testing from Windows Latest found that the built-in Windows 11 Weather app can consume more than 1.2GB of memory across multiple processes, all while it’s just sitting there. That’s over 1/8th the memory of Microsoft’s newest Surface. From a weather app…

What’s infuriating is that the app itself is effectively just a web app wrapped inside a Windows app, that also includes ads. Microsoft has truly fallen.

For comparison, Apple’s native weather app in MacOS takes up 1/5th as much memory with just a single process. I honestly still think that’s a bit much, but it clearly shows the difference between actual optimization and whatever Microsoft is doing.

Quick Bytes

Xbox Game Pass on Linux?!

The developers behind Heroic Launcher are working on an open-source project called Xodus that aims to reverse engineer the services needed to run modern Xbox PC titles on Linux. The project can already log into Xbox services, download packages and obtain licenses, although there’s obviously still a ton of work ahead.

One RTX 4090 Owner Got the RMA Jackpot

After reportedly going through four problematic RTX 4090s, Zotac ultimately upgraded their customer to an RTX 5090 for free. The strange part: the replacement reportedly received only a 30-day warranty, but I still think Zotac is a champ for doing this.

Crucial’s Exit is Creating Some Problems

One customer with a failing 48GB DDR5 kit reportedly received an initial reimbursement offer far below the kit’s current replacement cost before Micron eventually improved the offer. Micron had previously promised to continue supporting warranties after shutting down the Crucial consumer business.

Intel’s Big APU Gets Spotted

HWiNFO has added preliminary support for the rumored Razor Lake-AX, which is expected to be one of Intel’s upcoming high-end APUs. There still aren’t many confirmed specs yet, but software support showing up is usually a pretty good sign that the hardware is moving along behind the scenes. So this could actually be the real deal.

So yeah. A pretty weird week. If this keeps up, AM4 is going to outlive all of us.

Until next time, may your memory prices keep falling, your four-year-old CPU keep winning, and your weather app use less RAM than the game you’re playing.

I’m talking revenue increased 50 percent year over year to $11.536 billion. Data Center revenue reached $6.718 billion alone. That’s up 107 percent. Client revenue rose 23 percent to $3.062 billion. Everything looks great.

Until it doesn’t.

Let’s start with their client side of things. This is AMD’s consumer CPU business. Here, the company shipped 34% more Ryzen processors than it did a year ago. That’s a huge increase, and it helps explain how client revenue grew so much.

But there’s another number hiding underneath it.

AMD’s average selling price per processor actually fell by 6%. According to the company, most of that shipment growth came from Ryzen mobile processors, while desktop Ryzen sales actually declined. In other words, AMD sold a lot more CPUs, but the mix shifted toward cheaper products.

And the reason is pretty obvious: building a desktop PC has become ridiculously expensive. Even if AMD’s CPU prices stay reasonable, buyers still have to pay for memory, a motherboard, a graphics card, and everything else required for a new system. More people are simply deciding that upgrading isn’t worth it right now.

This brings us to gaming.

And this is where AMD’s record quarter really starts to look terrible.

Gaming revenue actually dropped to $779 million. That’s down a massive 31% from $1.1 billion just one year ago.

Now, AMD blamed most of the decline on their semi-custom business, meaning the chips that they sell for consoles. Specifically, that they’re reaching the late stage of the console lifecycle, but there’s a problem here. Because both Sony and Microsoft actually raised prices on their consoles due to the memory crisis. So that can easily account for at least some of that demand loss.

Plus, AMD’s CEO, Lisa Su, confirmed that their gaming GPU revenue declined year over year because higher industry-wide component costs pushed graphics-card prices higher and weighed on demand. CFO Jean Hu also revealed that Radeon shipments declined compared with the previous quarter.

To give you an idea of how bad all of this is, when Client and Gaming are combined, their revenue increased by 6%, but their operating income actually fell from $767 million to $582 million. Meaning their revenue is up, but profits are down.

AMD says that decline was primarily caused by higher operating expenses. But the result is still the same: this entire side of the company brought in more revenue and produced less operating income.

And the second half could be even worse.

AMD’s CFO, Jean Hu previously said that the company expects Gaming revenue to decline by more than 20 percent compared with the first half, largely because of higher memory and component costs.

Ultimately, AI is beginning to suck the life out of the consumer market.

And that’s the incredible irony behind these results. AMD is having the best quarter in its entire history because AI demand is exploding. While at the same time, its gaming business is collapsing because that same AI boom is making the hardware consumers need far too expensive.

The force making AMD richer is also making its consumer products harder to sell.

So yes, this was AMD’s greatest quarter ever.

But for PC gamers, it may be one of the company’s biggest warnings yet.

We Finally Have Competition

With that said, there may finally be one small piece of good news in this memory disaster.

According to a new report, HP, ASUS, and Acer have all qualified DRAM from China’s CXMT (the company I recently told you about) and they’ve already started using small amounts of it in certain laptops sold outside of the United States.

Now, the important thing here isn’t that a few laptops have Chinese memory inside of them.

It’s that CXMT has apparently passed the testing required by some of the biggest PC manufacturers in the world. Because companies like HP and ASUS aren’t going to put completely unproven memory inside their systems and hope for the best. That means CXMT is beginning to move from a mostly Chinese supplier into a legitimate fourth player in the global DRAM market. And that’s exactly what the industry needs right now.

Samsung, SK hynix, and Micron currently control almost the entire market. So when all three decide to prioritize more profitable AI memory, there aren’t many other places for PC companies to go.

CXMT won’t fix that overnight. Its memory reportedly isn’t even cheaper than chips from the major manufacturers yet, and the number of laptops using it is still extremely small.

But more suppliers eventually means more capacity and ultimately, more competition.

And CXMT is already expanding production while pushing its DDR5 memory to much higher speeds. So if the company continues improving, it could eventually take some pressure off the consumer market.

For now, this isn’t the end of the memory crisis.

But it may finally be the beginning of an actual solution.

Zen 6 Targets The Frames That Benchmarks Hide

Average FPS is a useful number. Unfortunately, it’s also very good at hiding a bad experience.

Believe it or not, a game can average well over 100+ frames per second and still feel terrible. That’s because average FPS doesn’t show how evenly those frames are delivered. If most frames arrive quickly but a few take much longer, you’ll feel a sudden hitch or micro-stutter even though the average still looks great. Those unusually slow frames are what drag down 1% lows, and improving them can make a game feel much smoother without dramatically increasing average FPS.

Well, according to new information shared by the well known Ryzen tuning developer: 1usmus, AMD is preparing several Zen 6 performance-management controls that could seriously help with those 1% lows.

CPPC Performance Priority

The first feature is called CPPC Performance Priority.

Basically, when a processor reaches its power or temperature limit, it has to start reducing performance somewhere. But instead of slowing every core down equally, Zen 6 could prioritize the cores that are actually running your game, while cutting performance on cores handling background tasks.

So if you’re gaming with Discord, a ton of browser tabs, RGB software, and several other programs open, the processor could take power away from those less important workloads and preserve the performance of the cores your game actually needs.

More Accurate Core Ranking

AMD is also reportedly improving how the operating system sees the maximum frequency of each individual core.

Because even inside the same processor, some cores naturally boost higher than others. Giving Windows more accurate information could help it place important gaming threads on the fastest cores more consistently.

It sounds like a small change, but putting the most important work on the wrong core can easily hurt performance and frame consistency.

Per-Core EPP Boost

Then there’s per-core EPP Boost, and this may be the most interesting feature of all.

During gaming, an important CPU thread may briefly stop while it waits for the GPU, another thread, or the next part of the game. The processor can interpret that tiny pause as the core no longer needing as much performance and lower its frequency.

Then, when the game suddenly needs that core again, the clock speed may not increase quickly enough.

That delay may barely affect average FPS, but it can create one unusually slow frame, which is exactly what hurts 1% lows and causes those noticeable stutters.

Per-core EPP Boost could temporarily keep a recently active core in a higher-performance state, allowing it to respond much faster when the work returns.

To give you an idea of how big that could be, a similar feature was recently tested on the original Steam Deck. Average performance barely changed, but 1 percent lows in Civilization VI reportedly improved by nearly 32%.

Background Memory-Bandwidth Limits

This one is big. According to the report, AMD may also let Zen 6 restrict how much memory bandwidth background tasks are allowed to use.

Basically, your browser, Discord, Windows services, game launchers, and the 47 tabs you definitely meant to close would no longer be able to fight your game for as many resources. That could help prevent sudden drops in performance when a background program decides it needs to do something at the worst possible moment.

And when you combine all of these features, Ryzen 10000 may not just improve average FPS. It could specifically target the slow frames that make an otherwise fast game feel terrible.

So on top of more cores, higher boost clocks, and a new process node, Zen 6 may have some major software-level gaming improvements as well.

Next generation is looking seriously impressive.

AMD’s Next Gen Gaming GPUs Are On The Way

AMD’s next Gen Radeon architecture just became a lot more real.

The company has started publishing Linux support for its new DCN 6 display engine, and the code directly connects it to GC 13.0.1. If you didn’t know, that’s part of the GFX13 family widely expected to power AMD’s next-generation gaming GPUs.

Now, we’ve seen references to GFX13 before, but most of those were buried inside compiler code. This is different. DCN 6 is an actual piece of the GPU responsible for handling display output, scaling, pixel formats, and how the final image is sent to your monitor.

The new patches include support for improved scaling of chroma-subsampled formats like 4:2:0 and 4:2:2, along with more control over chroma positioning and image quality. Basically, AMD is now adding support for specific hardware features instead of simply referencing the architecture by name.

AMD has already confirmed that it eventually plans to combine its gaming-focused RDNA architecture and data-center CDNA architecture under one unified design called UDNA. We still don’t know whether GFX13 will be the first full version of that architecture, but a leak from Moore’s Law Is Dead seems to claim that it is.

Either way, AMD’s next GPUs are clearly moving beyond the planning stage, and the company is already laying the software foundation needed before the hardware actually launches.

Quick Bytes

GPU Increases Spread Beyond the First U.S. Listings

High GPU prices are now spreading to other parts of the world.

In Japan, a distributor reportedly warned customers that certain Gigabyte graphics cards were increasing in price by 20 to 40 percent. Meanwhile, eight different RTX 50-series listings in South Korea jumped by roughly 30 percent on average in just a matter of days.

Now, this doesn’t mean every graphics card everywhere is suddenly 30 percent more expensive. These are specific models in specific markets. But that’s exactly what makes this important. What initially looked like a few isolated price hikes is now showing up across multiple countries, brands, and retailers.

AMD’s RX 9050 Makes the RTX 5050 Look Reasonable

One of the first published tests of AMD’s 8GB Radeon RX 9050 showed it trailing the RTX 5050 by roughly 9 to 30 percent across several traditional gaming and synthetic benchmarks. There were a few exceptions, but the overall picture wasn’t exactly flattering.

The card can handle 1080p gaming, and its 92-watt power target makes it an interesting upgrade for lower-power systems. Unfortunately, AMD’s $279 suggested price in select Asia-Pacific and Latin American markets puts it uncomfortably close to, or even above, faster alternatives.

One early review isn’t a final verdict. But if your new budget GPU makes the RTX 5050 look like the sensible purchase, you’re doing something bad.

16GB GPUs and 8 Core CPUs Take the Steam Lead

Valve’s July Hardware Survey shows 16GB graphics cards narrowly passing 8GB models, while 8 core CPUs also moved slightly ahead of 6 core chips.

Now, I’ve never loved putting too much weight on these surveys. Valve only asks a randomly selected group of Steam users, participation is optional, and the results can move around quite a bit from month to month. So this isn’t a perfect picture of every gaming PC in the world.

Still, it does give us a rough idea of where the market is heading. And right as 16GB GPUs and 8 core CPUs are becoming the new mainstream target, memory prices are exploding.

Excellent timing, as always.

That’s all the bytes for this week.

Until next time, may your 1 percent lows stay high, your memory prices stay low, and your next GPU never make the RTX 5050 look reasonable.

Over the past few days, new distributor pricing from MSI and Colorful showed increases across almost the entire RTX 50 series in China. Some cards jumped by as much as 20% in roughly one week.

At the time, there was still a chance that the increases would remain isolated to the Chinese market. Different regions have different inventory levels, demand, taxes, currencies, and retailer margins.

But now several premium RTX 50 cards have become far more expensive at Best Buy, including an RTX 5080 that costs more than Nvidia’s original MSRP for the RTX 5090.

So, unfortunately, the price increases are no longer limited to China.

They have reached the United States.

The First Warning Appeared in China

New distributor sheets from MSI and Colorful show higher prices across almost the entire RTX 50 lineup, ranging from the RTX 5050 to the RTX 5090 D V2.

MSI’s document is especially important because it includes both the previous and updated prices.

For example, the RTX 5080 Shadow and Ventus models reportedly increased from 9,999 yuan (approximately $1,478 USD) to 11,999 yuan (approximately $1,773 USD). That’s a whopping 20% increase!

The RTX 5080 Gaming Trio went from 10,999 (approximately $1,626 USD) to 12,999 yuan (approximately $1,921 USD), while MSI’s RTX 5070 Ti Shadow jumped from 7,699 (approximately $1,138 USD) to 9,199 yuan (approximately $1,360 USD). But that’s not all!

Even the more affordable cards were affected.

MSI’s RTX 5060 Shadow increased from 2,599 (approximately $384 USD) to 2,999 yuan (approximately $443 USD), and the RTX 5050 Gaming OC received the same 400-yuan adjustment.

Colorful’s document doesn’t show its previous prices, so we can’t calculate exactly how much each model increased during this specific change, but its current prices are still completely absurd compared with Nvidia’s official Chinese MSRPs.

Colorful’s least expensive RTX 5070 Ti now starts approximately 38% above MSRP, while some premium versions are nearly 59% higher. Its RTX 5080 models range from around 26% to 59% above MSRP.

The reported cause is exactly what you would guess: rising GDDR6 and GDDR7 memory costs.

Notices circulating among Chinese distributors reportedly suggest increases of approximately:

• 600 yuan for cards with 8GB of memory

• 900 yuan for cards with 12GB

• 1,200 yuan for cards with 16GB

That helps explain why the 16GB RTX 5070 Ti and RTX 5080 received some of MSI’s largest increases.

But MSI and Colorful are not necessarily where the increases began.

To understand what’s happening, we have to move one step further up the supply chain.

Nvidia Reportedly Raised the Cost of Building

GPU manufacturers don’t normally purchase a GPU from Nvidia and then shop around for the memory separately.

Instead, partner companies like ASUS, MSI, Gigabyte, Zotac, etc. typically purchase a complete kit containing both the graphics processor itself and the GDDR6 or GDDR7 memory needed to build the card. Those companies then design the PCB, cooler, power-delivery system, and finished graphics card that ultimately reaches retailers.

Well, according to a report from BenchLife, Nvidia recently informed its board partners that the price of those GPU and memory kits was increasing.

Nvidia had reportedly already adjusted kit pricing for the RTX 5090 and China’s RTX 5090 D V2 in May.

The newer increase apparently extends to additional GeForce products using both GDDR7 and the older GDDR6 standard. The exact models and size of the adjustments were not disclosed, but given the increases from MSI and Colorful, I would assume it’s a lot.

Board partners are also reportedly paying more for coolers, PCBs, and packaging, although those increases are said to be smaller than the additional memory costs.

Nvidia and its partners have not publicly announced retail price changes directly connected to the reported adjustment. But only days after the report appeared, the new MSI and Colorful distributor sheets showed widespread increases across China.

And now the same thing is beginning to appear in the United States.

An RTX 5080 Just Passed the 5090’s MSRP

According to VideoCardz, Best Buy has raised the prices of several premium RTX 50-series GPUs.

The ASUS ROG Astral RTX 5080 OC is now listed for an unreal $2,099.99! Up from its previous Best Buy price of $1,899.99.

That makes this particular RTX 5080 $100 more expensive than Nvidia’s original $1,999 MSRP for the RTX 5090.

Yes, an RTX 5080 now costs more than the RTX 5090 was originally meant to cost.

Apparently, calling it “Astral” meant the price also needed to leave Earth’s atmosphere.

When ASUS first launched the card, its official price was $1,499.99. That means its Best Buy price has increased by $600, or 40%, since launch. Several premium RTX 5090 models have also moved higher.

The ASUS ROG Astral RTX 5090 is now listed for $4,829.99, while MSI’s RTX 5090 Suprim Liquid SOC has reached $4,599.99. For that kind of money, the GPU should at least come with the rest of the computer.

With that said, there is an important caveat.

These increases have not affected every RTX 50 card at Best Buy. Other RTX 5080 models remain significantly cheaper, meaning this currently appears to be a price adjustment affecting certain premium cards rather than a uniform increase across the entire lineup.

But the sequence of events is difficult to ignore.

First, Nvidia reportedly increases the cost of GPU and memory kits. Then Chinese distributors raised prices across almost the entire RTX 50 series. And only days later, several premium cards at one of America’s largest electronics retailers become even more expensive.

The U.S. increases are not yet as widespread as the new distributor pricing in China. But they are now here.

And unless memory costs begin moving in the opposite direction, these premium cards may only be the beginning.

Valve May Have Found the Breaking Point

There is at least one force powerful enough to prevent companies from raising prices forever…

Customers refusing to pay.

In May, Valve increased the price of their 512GB Steam Deck OLED from $549 to $789, and the 1TB model jumped from $649 to $949. That’s a massive 44% increase for the 512GB version and a 46% increase for the 1TB model.

Valve blamed the changes on rising memory and storage costs. And I have no doubt that’s true, but according to a new analysis from Boiling Steam, customers may have responded exactly how you would expect.

Before the increase, the Steam Deck was estimated to sell between 11,000 and 18,000 units per week. Afterward, estimated weekly sales fell to somewhere between approximately 1,400 and 3,000 units.

That represents an unbelievable decline of as much as 82%!

With that said, these are not official figures from Valve. The company doesn’t publish weekly Steam Deck unit sales, so the analysis uses the device’s position on Steam’s best sellers chart, estimated revenue associated with each position, and the Steam Deck’s higher average selling price.

Meaning there are multiple assumptions involved, so the exact drop off could be smaller, or potentially even bigger. Regardless, Steam’s own rankings support the broader conclusion that demand fell considerably.

Before the increase, the Steam Deck generally remained around fourth or fifth place on Steam’s weekly chart. Afterward, it fell into approximately the tenth to fifteenth place range.

That is what people mean when they say companies charge what the market can bear.

A company cannot simply select any price and expect customers to continue buying the same number of products. It can keep raising prices only until the additional profit earned on each device is no longer enough to offset all the customers who stop buying.

Valve increased its prices largely because its own costs were rising. The company was attempting to protect its margins, but if the decline in demand is anywhere close to this estimate, the adjustment may have hurt Valve more than it helped.

The Market Has Only Looked Like This Once Before

One of the biggest stories of the year I would say, is that the Shiller price-to-earnings ratio, or CAPE ratio, reached approximately 41 during July.

That may sound like an obscure financial statistic, but the underlying idea is fairly simple. A normal price-to-earnings ratio compares the current price of a stock with the company’s recent earnings.

The problem is that earnings can rise or collapse during a single year because of a recession, temporary shortage, tax change, or any other unusual event.

The Shiller P/E attempts to smooth out that noise.

Instead of using only one year of profits, it compares the value of the S&P 500 with the average inflation-adjusted earnings of those companies over the previous ten years.

Historically, the ratio’s average is around 17.4.

Its all-time record was approximately 44.2 in December 1999, near the peak of the dot-com bubble.

There have only been two modern periods when the ratio remained around or above 40 for a meaningful amount of time.

The first was the dot-com bubble.

The second is now.

That does not mean the market will collapse tomorrow.

Valuation measurements can show that stocks are historically expensive, but they are generally terrible at predicting the exact moment when investors will stop paying those prices.

There are also legitimate differences between the companies driving today’s AI boom and many of the businesses receiving money during the dot-com era.

Microsoft, Nvidia, Amazon, Alphabet, and Meta generate enormous revenue, profit, and cash flow. They aren’t three employees, a website, and a business plan written on a napkin. But profitable companies can still become overpriced.

The internet genuinely transformed the global economy. It changed how people communicate, shop, work, access information, and watch entertainment.

That did not prevent the dot-com bubble from collapsing.

A revolutionary technology can be completely real while the amount investors are willing to pay for it is still completely insane.

That is the risk facing AI.

AMD’s Next Two CPU Generations Could Both Be Massive

AMD recently revealed a processor roadmap extending through Zen 8, and the company appears to be done saving its largest manufacturing improvements for every other generation.

Its Zen 6-based EPYC Venice processors use TSMC’s advanced 2nm technology. Current Zen 5 compute dies are manufactured using a 4nm process, meaning AMD is effectively skipping the traditional 3-nanometer step for its next major CPU architecture.

Venice scales to as many as 256 cores and 512 threads, supports up to 16 channels of DDR5 memory, and adds PCIe 6.0 connectivity.

AMD has already begun ramping production, describing Venice as the first high-performance computing product to enter production using TSMC’s advanced 2-nanometer process. But here’s the thing: AMD’s also planning another manufacturing transition immediately afterward.

The company’s Zen 7-based Florence processors are scheduled for 2028 and will use what AMD calls a “next-generation process node.” AMD has not officially named that node, but a separate report claims the Zen 7 compute die, codenamed Grimlock, could use TSMC’s A14 technology, but that remains unconfirmed.

AMD has also confirmed that development has begun on Zen 8, with the EPYC Ravenna family planned for approximately 2030. That means AMD could move from 4nm with Zen 5, to 2nm with Zen 6, and then make another process-node transition with Zen 7 just two years later.

The company is apparently done taking its foot off the gas.

Nvidia’s First Consumer CPU Was Just Tested

You’re never going to believe this, but someone “found” an alleged prototype Microsoft Surface Laptop Ultra with Nvidia’s upcoming RTX Spark inside. And he actually spent a month testing it!

The system reportedly uses Nvidia’s top N1X configuration, combining:

• A 20-core Arm processor

• A Blackwell GPU with 6,144 CUDA cores

• 24GB of unified memory

As far as performance, the CPU didn’t do terrible.

In Cinebench 2026, the processor scored 5,771 points in the multicore test and 540 points in single-core. That places it in roughly the same range as Apple’s 14-core M3 Max. And that’s not bad for Nvidia’s first major attempt at a mainstream Windows consumer processor.

Unfortunately, gaming was a completely different story.

According to the tester, GPU clock speeds repeatedly fluctuated between approximately 1.5GHz and 2.3GHz, producing major stutters every few seconds. Increasing the system’s power limit made almost no measurable difference, while certain workloads and games crashed entirely.

Temperatures were also brutal.

The CPU reportedly reached between 98 and 100 degrees Celsius, while the GPU generally operated between 80 and 88 degrees. Desktop idle consumption reached as high as 23 watts in the system’s highest-performance mode.

That all sounds terrible, but this is an early engineering sample running unfinished software and drivers. The results should not be treated as representative of the final retail product.

In fact, the most interesting takeaway may be that Nvidia’s underlying hardware already appears surprisingly capable. The company seems to have created a processor capable of competing with some of the most powerful laptop CPUs on the market.

Now it needs to make Windows, drivers, power management, and existing games work properly on it.

That may prove to be the much harder part.

QUICK BYTES

Intel May Be Building a Real Desktop APU

A preliminary Intel Nova Lake desktop configuration reportedly combines four P-cores, eight E-cores, four low-power E-cores, and 12 Xe3P graphics cores.

That would be a much larger integrated GPU than Intel normally includes inside its socketed desktop processors.

If the configuration reaches production, it could give Intel a proper desktop competitor to AMD’s Ryzen G-series for compact systems and lower-cost gaming PCs that do not need a separate graphics card.

The specifications remain unconfirmed, and Intel has not announced clocks, memory support, pricing, or performance targets.

Someone Built a Fanless RTX 4060 With 2.5 Kilograms of Aluminum

A PC builder has created a completely fanless RTX 4060 by attaching an enormous 2.5-kilogram block of aluminum directly to the card.

The cooler relies entirely on natural convection, allowing warm air to rise through the widely spaced aluminum fins without requiring a fan.

The design is obviously far too large and heavy for a normal PC case, but the relatively efficient 115-watt RTX 4060 makes it one of the few modern graphics cards capable of making the experiment work.

It is silent, technically functional, and only slightly heavier than a small child.

Seagate Is Preparing 50TB Hard Drives

Seagate plans to begin qualification shipments of its first 50TB-class hard drives in late 2027, with wider shipments expected afterward.

The drives will use Seagate’s next-generation Mozaic 5 HAMR platform, with platters capable of storing more than 5TB each. The company’s longer-term roadmap includes approximately 60TB drives around 2029 or 2030, followed by 80TB and eventually 100TB models.

Naturally, much of the demand driving these enormous capacities is coming from AI and cloud companies.

Because apparently every part of the PC industry eventually leads back to AI buying all the hardware.

That’s all the bytes for this week.

Until next time, may your frame rates stay high and your component prices return to Earth.

A $280 Gaming CPU in a $400 RAM World

AMD’s new Ryzen 7 7700X3D perfectly illustrates how strange the market has become.

When AMD announced the processor at Computex, I wasn’t exactly impressed. The 7700X3D is essentially a slightly slower Ryzen 7 7800X3D. Meaning it comes with the same eight Zen 4 cores and 3DV-cache. It just lowers the clocks a bit.

At its official $329 price, it was an absolute no from me.

The 7800X3D is currently $339 at Micro Center, so there’s no world that makes a $329 7700X3D make sense to me.

Then launch day arrived.

Newegg immediately offered a $49 promotional discount, dropping the processor to $279.99 with code PKC337. As of this writing, the retailer is still advertising that promotion. At $280, the entire argument changes. You’re no longer saving $10 or $20—you’re potentially saving around $60 or more compared with readily available retail 7800X3D listings.

That turns the 7700X3D into a genuinely strong gaming CPU for anyone who already owns an AM5 motherboard.

The words “already owns” are important, though.

A cheap processor does not automatically create a cheap computer. DDR5 prices have jumped so high that the cost of memory can wipe out practically everything you saved on the CPU. Add a motherboard, RAM and storage, and suddenly that $280 gaming chip is sitting in the middle of a system that costs far more than an equivalent build would have a few years ago.

There’s also a strange availability problem. The 7700X3D is currently exclusive to Newegg in North America and reportedly won’t appear at additional U.S. or Canadian retailers until at least the fourth quarter. It’s being sold more broadly in other regions, but North American buyers currently have one official option.

So yes, the 7700X3D is now a great deal.

It’s just a great deal inside a hardware market where practically everything surrounding it has become painfully expensive.

Nvidia Built the GPUs. Memory Said No.

That brings us to Nvidia’s RTX 50 SUPER series.

According to a new report, at least one Nvidia board partner has already received physical RTX 50 SUPER GPUs. These reportedly aren’t preliminary specifications or extremely early prototypes, but the final product. The hardware exists, but Nvidia has instructed their partners to hold the cards while the company decides what to do about pricing.

The problem is reportedly the new 3GB GDDR7 memory modules.

Using those higher-density chips will let the 5070 SUPER go from 12GB to 18GB, and RTX 5070 Ti SUPER and RTX 5080 SUPER could jump from 16GB to 24GB.

That would address one of the biggest complaints surrounding Nvidia’s current generation: expensive graphics cards shipping with memory capacities that feel increasingly inadequate.

Unfortunately, the 3GB chips are reportedly absurdly expensive.

The source claimed one 3GB GDDR7 module currently costs somewhere around $60 to $70, compared with roughly $20 for a standard 2GB module. Nvidia and its partners almost certainly negotiate different volume pricing, so those figures shouldn’t be treated as the exact bill of materials. But the larger chips would still be considerably more expensive. Multiply that across six modules on an RTX 5070 SUPER or eight modules on a 5070 Ti SUPER or 5080 SUPER, and Nvidia is left with two options:

Charge customers substantially more or accept lower margins. As you would guess, Nvidia doesn’t like either of those.

The same memory problem is also reportedly holding back a 9GB version of the RTX 5050 that would use three of those higher-density GDDR7 chips. This is where the hardware market starts looking completely backwards.

Nvidia may have finished graphics cards sitting inside its partners’ facilities, yet it still can’t find a price that makes sense. The GPU itself isn’t the bottleneck. The memory surrounding it is.

And it probably won’t stop with graphics cards.

The Temperature Nvidia Didn’t Want You to See

The RTX 50 series is also generating controversy for a completely different reason. When these GPUs launched, normal monitoring software could no longer display the chip’s hotspot temperature. Users could see the general GPU reading, but not the temperature of the hottest individual section of the silicon.

That matters because an average temperature can look perfectly acceptable while a specific section of the chip is running dangerously hot.

Last time I discussed a Brazilian repair specialist, Paulo Gomes, who recently demonstrated exactly that a hotspot approaching 107 degrees thanks to dried thermal material and an uneven cooler installation. Once corrected, temperatures and performance improved.

Developers have since restored RTX 50 hotspot monitoring through HWMonitor, HWiNFO, AIDA64 and a community plugin for MSI Afterburner.

And owners are already discovering questionable results.

One Colorful RTX 5080 owner reportedly measured a hotspot temperature close to 100 degrees. There may have been a measurement error depending on which version of HWMonitor was used, but Colorful’s response was arguably more important than the exact reading.

The company recommended checking airflow and dust buildup first. However, if the hotspot remains at or above 95 degrees Celsius under otherwise good cooling conditions, Colorful recommends submitting the card for an after-sales inspection to check for a cooler defect or another hardware issue.

That doesn’t mean every RTX 50 GPU has a cooling problem. It doesn’t even mean a chip briefly touching 95 degrees will instantly fail.

But sustained high temperatures can accelerate degradation, reduce boost behavior and reveal poor cooler contact that users otherwise have no way to diagnose.

There was never a good consumer-facing reason to hide this reading.

When a $1,000+ graphics card has a cooling defect, the solution should not be hoping the average temperature hides it convincingly enough.

Meanwhile, Next-Generation CPUs Look Insane!

While current hardware struggles with pricing and availability, next-generation CPUs are moving in the exact opposite direction.

A 10-core AMD Medusa Point engineering sample recently appeared in Geekbench with a single-core score of 3,329 and a multi-core score of 16,555. The internal result indicated a peak clock of around 5.37GHz despite the early listing displaying much lower frequencies.

Compared with average results for the 12-core Ryzen AI 9 HX 370, the engineering sample was approximately 28% faster in single-core performance and 24% faster in multi-core—even though the new chip has two fewer cores.

Compared with the existing 10-core Ryzen AI 9 365, the differences grow to around 35% in single-core performance and 33% in multi-core.

Even more impressively, the same engineering-sample identifier had appeared previously with scores of 3,174 and 15,092. The newer run was approximately 5% faster in single-core and nearly 10% faster in multi-core. YET it’s the same silicon revision! Meaning this bad boy gained this much without new silicon, so the final retail version could be way faster.

With that said, as always, leaked Geekbench results should be treated carefully. We don’t know the system’s power limits, cooling configuration or whether these results represent final retail behavior.

But early samples normally aren’t supposed to embarrass the existing lineup this badly.

And it doesn’t stop there, as Intel is preparing an equally aggressive response with their next Gen Nova Lake.

Leaked scheduling information suggests the first 28-core desktop processors could arrive between January and March of 2027, followed by unlocked K-series models around March or April. Smaller 16-core and 8-core models would follow, while Intel’s enormous 52-core flagship may not launch until sometime between May and September. The lineup is also expected to launch under Core Ultra Series 400 branding.

In other words, AMD and Intel are preparing some of the most ambitious consumer processors we’ve ever seen.

The only question is what the rest of the computer will cost by the time they arrive.

Nvidia Wants to Build the Entire PC

And the CPUs don’t stop there, as Nvidia is preparing to enter the CPU battle as well.

The company recently released GeForce driver 616.00, its first native GeForce developer-preview driver created specifically for Windows 11 on Arm and the upcoming RTX Spark platform.

Files inside the driver confirm two high-end RTX Spark configurations.

The full model contains 6,144 CUDA cores paired with a 20-core Arm CPU, while a cut-down version uses 5,120 CUDA cores and an 18-core CPU. These match the previously leaked N1X 675 and N1X 650 configurations. Both can reportedly support as much as 128GB of unified LPDDR5X memory shared between the CPU and GPU.

That could be extremely interesting for gaming, content creation and local AI workloads, but it also creates an enormous software challenge.

Windows on Arm has improved significantly, but Nvidia needs mature graphics drivers, reliable x86 translation and game developers willing to support the platform. Anti-cheat software remains an especially important obstacle because many competitive games still rely on low-level components that don’t automatically translate across CPU architectures.

The hardware may be fast enough.

Whether everything runs properly is the part Nvidia still has to prove.

The $1.65 Trillion Bet Behind Everything

All of these stories ultimately lead back to the same place: the unprecedented amount of money being spent on AI infrastructure.

A recent report estimates that Alphabet, Amazon, Meta, Microsoft and Oracle collectively have approximately $1.65 trillion in off-balance-sheet commitments. That’s roughly 122% more than the debt they already have on their balance-sheet.

Calling it “hidden debt” makes it sound more sinister than it necessarily is.

These companies aren’t secretly fabricating financial statements. Much of the total reportedly consists of long-term data-center leases, power agreements and compute contracts that have been signed, but haven’t taken effect yet. Under normal accounting practices, those obligations may not appear as traditional debt until the projects become operational.

Once they do, the companies have committed to paying for that capacity whether demand ultimately matches their expectations or not.

That is where the real risk begins.

Reuters estimates that the five hyperscalers’ annual capital expenditures could increase by approximately $534 billion by 2027, while their annual operating cash flow rises by only around $340 billion.

That works out to roughly $1.57 in additional investment for every $1 of additional operating cash generated. Oracle’s free cash flow has already turned negative, Amazon and Microsoft’s have fallen sharply.

To be clear, this does not mean that AI demand is fake or that AI is useless.

The internet permanently changed the world, and investors still created a massive dot-com bubble around it. A transformative technology and a financial bubble can exist at the same time.

The issue is that every major technology company is terrified of falling behind.

Microsoft spends more, so Google spends more. Google spends more, so Meta spends more. Nvidia develops a larger AI system, so AMD develops a competing rack. Cloud providers reserve GPUs, CPUs, power and data-center capacity years in advance because waiting could mean losing customers to a competitor.

That creates demand based not only on what companies need today, but on what they fear their rivals may have tomorrow. And as long as that spending continues, consumer memory and storage prices may remain painfully high. Manufacturers have every incentive to prioritize HBM, server DRAM, enterprise SSDs and other products that generate larger margins than ordinary desktop components.

But markets like this often overshoot in both directions.

If AI spending slows just as memory companies complete new production capacity, the shortage could very quickly become an oversupply. Companies that spent years producing every chip they could will suddenly begin competing to unload inventory.

That is the scenario where RAM and SSD prices don’t merely decline. They collapse.

It isn’t guaranteed. AI demand may continue growing quickly enough to absorb practically everything manufacturers can produce. But the longer companies expand based on aggressive projections and competitive fear, the more violent the correction could become if those projections disappoint.

For consumers, that leaves two very different possible futures.

Either memory remains expensive for years because AI demand continues consuming practically everything—or this becomes one of the largest boom-and-bust cycles the semiconductor industry has ever experienced.

Right now, the industry is betting $1.65 trillion that it will be the first one.

What do you think is more likely: memory prices stay painfully high for years, or the AI buildout eventually causing a massive oversupply crash?

Leave a comment

QUICK BYTES

Framework’s RAM Pricing Nearly Doubled Overnight

Framework increased the price of a 32GB LPCAMM2 module for its Laptop 13 Pro from $439 to $800. The 64GB option jumped from $849 to an incredible $1,600.

Framework’s CEO said absorbing the supplier increases would put the company’s ability to operate at financial risk.

At that price, the memory upgrade is beginning to look less like a laptop component and more like a lightly used car.

Even DDR2 Isn’t Safe

TrendForce estimates that DDR2 contract prices have increased approximately 55% to 60% during the second quarter of 2026 and could rise by another 35% to 40% in the third quarter.

Yes, you heard that right… DDR2.

The memory standard used by computers old enough to have strong opinions about Windows Vista is experiencing a supply crisis in 2026.

Server CPUs Are Getting More Expensive Too

Intel and AMD are reportedly entering longer-term supply agreements with Chinese customers as AI infrastructure demand creates shortages extending beyond GPUs and memory.

Prices for some server processors in China have reportedly increased by more than 40% since the beginning of the year, while lead times on certain Intel products have reached as long as six months.

Xbox Games Are Finally Coming to PC—Properly

Microsoft announced Xbox Backward Compatibility for PC, allowing select classic Xbox games to run on Windows PCs and handhelds.

The first releases include Blinx: The Time Sweeper, Conker: Live & Reloaded, Crimson Skies: High Road to Revenge and Fuzion Frenzy. Digital owners won’t need to repurchase supported titles, and the games will also receive Xbox Play Anywhere and cloud support.

It’s only four games to start, but officially bringing the console backward-compatibility library to PC could eventually become a very big deal. Then again, this could be the precursor to Microsoft announcing an end to discs just like Sony. Fingers crossed that isn’t the case.

Share

That’s all the bytes for this week.

Until next time, may your GPU hotspot remain below 95 degrees and your RAM kit cost less than the rest of your computer.

To thank everyone for supporting the launch of Meld Bytes, I’m giving away a G.SKILL Trident Z5 Neo RGB 32GB (2×16GB) DDR5-6400 memory kit memory kit to one lucky subscriber.

The memory kit was previously used for testing, but it’s fully functional.

How to enter

It’s simple:

Be an active Meld Bytes email subscriber at 11:59:59 PM, Eastern Time on 08/22/2026.

Already subscribed? You’re automatically entered.

Not subscribed yet? Enter your email at MeldBytes.com before the giveaway closes.

There are no forms to complete, no purchase is necessary, and both free and paid subscribers are eligible.

Choosing the winner

After the giveaway closes, I’ll randomly select one eligible subscriber.

The winner will be contacted at the email address associated with their Meld Bytes subscription and will have 72 hours to respond. If the selected winner does not respond within that period, another winner may be selected.

Important: I will never ask the winner to pay shipping fees, provide banking information, purchase anything, or contact me through Telegram or WhatsApp.

Good luck, and thank you for being a part of Meld Bytes!

Official Rules

1. Sponsor

This giveaway is sponsored solely by Gamer Meld LLC, referred to below as the “Sponsor.” Questions regarding the giveaway may be directed to admin@gamermeld.com.

2. Eligibility

By participating, entrants agree to be bound by these Official Rules and the decisions of the Sponsor, which are final and binding regarding this giveaway. A potential winner may be required to confirm acceptance of these Official Rules before receiving the prize.

The giveaway is open to legal residents of the 50 United States and the District of Columbia who are at least 18 years old at the time of entry.

Employees, contractors, and immediate family members of the Sponsor are not eligible. Void where prohibited by law.

3. Giveaway period

The giveaway begins at 5:00 p.m. Eastern Time on July 22, 2026, and ends at 11:59:59 p.m. Eastern Time on August 23, 2026.

4. How to enter

To enter, an individual must be an active Meld Bytes email subscriber when the giveaway closes.

Existing subscribers are automatically eligible. New subscribers may enter for free through MeldBytes.com.

No purchase is necessary. A purchase or paid subscription will not improve the odds of winning. Limit one entry per person. Duplicate, fraudulent, automated, or otherwise invalid entries may be disqualified.

5. Prize

One winner will receive:

32GB Trident Z5 NEO RGB 6400 DDR5 memory kit

The prize was previously used for testing and is being provided in fully functional condition. Approximate current retail or fair-market value: $400.

No cash substitution is available. The Sponsor may substitute a prize of comparable value if necessary.

6. Winner selection and odds

The winner will be randomly selected from all eligible active Meld Bytes subscribers on or about August 23, 2026. Odds of winning depend on the total number of eligible entries received during the Giveaway Period.

7. Winner notification

The potential winner will be contacted using the email address associated with their Meld Bytes subscription.

The potential winner must respond within 72 hours and may be required to confirm eligibility and provide a valid shipping address. If the potential winner cannot be contacted, does not respond in time, or is found to be ineligible, another winner may be selected.

The Sponsor will never require payment to claim the prize.

8. Shipping and taxes

The Sponsor will cover standard shipping to an eligible address within the 50 United States and the District of Columbia.

The winner is responsible for any taxes or other expenses associated with receiving or using the prize.

9. Privacy

Subscriber email addresses will be used to deliver Meld Bytes, administer the giveaway, verify eligibility, and contact the winner.

Personal information collected for the giveaway will not be sold or used for unrelated giveaway marketing. The winner’s shipping information will be used only to deliver the prize, except where disclosure is required by law.

10. Third-party disclaimer

YouTube, Substack, and G.Skill (manufacturer of the prize) are not sponsors, endorsers, or administrators of this giveaway, and are not responsible for its administration, prize, or selection of the winner. By entering, participants fully release YouTube, Substack, and G.Skill from any and all liability related to the giveaway. Participants must comply with all applicable platform community guidelines.

11. General conditions

The Sponsor reserves the right to disqualify entrants who attempt to manipulate the giveaway, submit fraudulent entries, or violate these rules.

The Sponsor may modify, suspend, or cancel the giveaway if fraud, technical problems, legal requirements, or circumstances beyond the Sponsor’s control prevent it from being conducted fairly.

The prize is awarded “as-is” with no warranty or guarantee, either express or implied, by the Sponsor.

But several developments this week suggest the supply situation could finally begin to improve.

The biggest comes from China’s CXMT, which is reportedly on track to reach approximately 350,000 DRAM wafer starts per month by the end of the year. That would put it only around 25,000 behind Micron’s projected capacity. The Chinese government is also reportedly encouraging CXMT to share its memory technology with three additional domestic manufacturers, all of which either have production capacity already or are preparing to bring it online.

We do not yet know how that capacity will be divided between DDR4, DDR5, mobile memory, server products and HBM. But almost any meaningful increase in supply could help.

Even additional HBM production matters because the current problem is not simply that manufacturers cannot make consumer memory. It is that HBM is so profitable that companies have been prioritizing it over consumer products. More competition and supply in the HBM market could eventually reduce the incentive to abandon everything else.

There may also be some help coming from a new memory standard called SPHBM4.

Traditional HBM4 uses an extremely wide 2,048-bit interface, which provides excellent bandwidth and efficiency but requires a massive number of physical connections between the memory and processor. That makes it heavily dependent on expensive advanced-packaging technologies such as silicon interposers and TSMC’s CoWoS.

SPHBM4 cuts the external interface to 512 bits and compensates by running those connections much faster. The goal is to preserve HBM4-class bandwidth while reducing the packaging complexity that has become one of the industry’s biggest bottlenecks.

Of course, this does not guarantee cheap RAM. AI companies may simply absorb every additional chip that manufacturers can produce. SPHBM4 also comes with tradeoffs involving signaling complexity and potentially lower efficiency.

Still, between China’s huge production expansion and an attempt to loosen the packaging bottleneck limiting HBM, the memory market may finally be attacking the shortage from more than one direction.

The Byte That Matters

More memory production does not guarantee that DDR5 prices collapse tomorrow. But between China expanding conventional DRAM output and the industry attacking the HBM packaging bottleneck, this is finally a potential supply solution—not just another attempt to sell consumers five-year-old hardware as the answer.

AMD’s Going Up to 8× Frame Generation?!

Hidden settings inside AMD’s Radeon drivers suggest the company is experimenting with multi-frame generation ratios reaching as high as 8×.

The options were discovered through RadeonTuner and include settings for ratios between 1X and 8X. At the highest setting, the GPU could theoretically generate seven artificial frames for every one traditionally rendered frame.

The feature does not currently work in games, however. AMD sometimes adds driver-profile options before the underlying functionality is ready, so this appears to be evidence of ongoing development rather than a usable feature.

AMD has already confirmed that its next-generation FSR technology will support multi-frame generation, making it likely that some version of this is coming.

The question is whether 8× frame generation is actually useful.

Generated frames can make motion appear smoother, but they do not process new player input. A game rendering natively at 20 FPS may be displayed as something resembling 160 FPS, but it will still feel like a 20 FPS game. To make matters worse, the frame-generation process can add some additional latency on top of that.

The Byte That Matters

AMD could leapfrog Nvidia at its own game. With that said, 8X frame gen really sounds more like a gimmick—and 6X frame gen does too, for that matter too. It can be visually nice if you have at least 60 FPS, but until the technology can accept input, real frames are still the way to go.

Nvidia’s Hotspot Reading Is Back

Since the initial release of Nvidia’s RTX 50-series cards, the company removed public access to the GPU’s hotspot temperature reading, including through their API used by programs such as HWiNFO and GPU-Z.

That was concerning because the hotspot reading measures the hottest area of the GPU die, which can be dramatically warmer than the average temperature shown to users.

A Brazilian GPU repair technician demonstrated why that matters while diagnosing an RTX 5070 Ti. The card showed a fairly normal average temperature of around 68°C, while Nvidia’s internal diagnostic software reported hotspot temperatures reaching approximately 106–107°C. The underlying issue appeared to involve poor cooler contact and dried thermal material.

Now, both HWMonitor and HWiNFO have restored RTX 50 hotspot monitoring—but not because Nvidia officially brought it back.

According to the developer behind MSI Afterburner, the programs are reportedly accessing the information directly through the GPU’s hardware registers rather than Nvidia’s supported NVAPI interface. MSI Afterburner cannot use the same method because it is officially partnered with Nvidia, although the reading may be accessible through the RivaTuner project that a lot of people download with Afterburner.

So Nvidia didn’t actually reverse their decision.

Developers just found a back door.

The First Zen 6 Benchmark Looks Ridiculous

An early benchmark for AMD’s upcoming Zen 6-based Medusa Point mobile processor was recently found in Geekbench.

The engineering sample reportedly includes 10 cores and 20 threads, divided between four traditional cores and six efficiency cores. Despite Geekbench reporting a weirdly low maximum frequency, the chip apparently outperformed the 10-core Ryzen AI 9 365 by an unreal 28.5% in single-core performance and 21.6% in multi-core performance.

The wild thing about these results is that the CPU appears to be an early engineering sample, meaning there’s a ton of potential for even more performance before release.

Ultimately, Zen 6 continues to look like a massive architectural leap—the likes of which we haven’t seen in nearly a decade.

Quick Bytes

RTX 50 Super Looks To Be Coming This Time

Seasonic has added the rumored RTX 5080 Super to its power-supply calculator after previously listing the 5070 Super and 5070 Ti Super. The listing gives the 5080 Super a 415-watt rating, compared with 360 watts for the regular model. Rumors also point to 24GB of GDDR7 memory, so a 50% increase over the existing RTX 5080.

Shader Stutter May Finally Become Less Annoying

AMD and Microsoft have expanded Advanced Shader Delivery to every RDNA generation, including RDNA 1 and RDNA 2. The system downloads shaders compiled specifically for the user’s hardware, potentially reducing initial loading times and in-game shader-compilation stutter. AMD claims load-time reductions reaching 95%, although support remains limited to compatible games through the Xbox PC app for now.

AMD Is Bringing Back Some Ancient CPUs

AMD quietly introduced three entry-level processors based on Zen+ and Zen 2, including the eight-core, 16-thread Ryzen 7 4700LE. The chips are primarily aimed at OEM systems and offer manufacturers cheaper DDR4-compatible options while newer memory remains expensive. It is not exactly technological progress, but it is at least another way to avoid paying current DDR5 prices.

AMD Has Decided to Pick a Fight With Apple

AMD launched a direct comparison against Apple’s budget MacBook Neo, emphasizing Windows gaming support, touchscreen options, storage and connectivity. AMD notes that 15 of the top 20 PC games do not run natively on the MacBook, although the Ryzen 5 system used in AMD’s comparison is not exactly a gaming powerhouse either.

Thank you for being a part of the first official Meld Bytes issue! As always, have a great day.

Hey, I’m Gamer Meld, and this is Meld Bytes.

For years, I’ve covered the biggest PC hardware and gaming technology stories through my YouTube channel: Gamer Meld. Now I’m expanding that coverage with Meld Bytes.

Every Thursday, I’ll bring together the most important news from across the industry—including new hardware, major company announcements, credible leaks and rumors, pricing changes, and updates to stories that are still developing.

Some of it may be news I’ve already discussed in a video, but Meld Bytes will give me room to cover even more. The goal is to create a more complete recap of everything worth knowing, including stories and details that I may not have had time to cover elsewhere.

It will also give me more space to share my own opinions.

Instead of only explaining what happened, I’ll be able to dig further into what I think it means, whether the hype is justified, where the industry may be headed, and how these developments could affect your next PC build or upgrade.

Meld Bytes isn’t replacing Gamer Meld. It’s an expansion of it: more coverage, more context, and more of my honest take on what’s happening.

Thanks for being here at the beginning.