jaxa

AMD can absolutely take the crown, but they are hesitant to do it. This could be the most obvious opening ever, with Nvidia delaying RTX 50xx Super cards but still probably planning on using them as a stopgap to stretch out the time until 60xx cards. Nvidia also has CPU overhead problems, see below. On the other hand, if GDDR7 remains expensive, it's a really bad time to do anything for consumers. All bets are off with respect to launch timing, given AMD launching Zen 6 Epyc several months before desktop. No reason that chiplet-based RDNA5 can't debut in laptop APUs many months before any desktop GPUs using the same silicon.

*384 I think you would need a real expert and/or a retrospective after a lot of time has passed to determine if it matters. But think of it this way. We the AMDumbs people want AMD to create a CUDA-like ecosystem where you can pick up any consumer card and get decent AI/HPC support, right? That's seemingly the point of UDNA, maybe also to reduce redundant work and driver issues. We already know that what's being called RDNA5 is going to be a long-lasting, important generation, because it will be the basis for next-gen consoles from Xbox and PlayStation. Xbox without any modification to RDNA5 at all, as it's rumored to be using the mid-range chiplet (the one closer to the likes of the 9070 XT, 5080, or 4090, not the 6090 killer) shared with desktop GPUs and other products. PS6 may not technically be pure RDNA5 since Sony likes to tweak it or even do AMD's work for them. So if the grand "UDNA-ification" isn't happening during the so-called RDNA5 generation, that seems like a missed opportunity, as that's a very important architecture that will be sticking around on the market in at least console form for basically the 2030s. Does RDNA5 == UDNA1 in everything but name? Maybe. But it would be better for it to be everything in everything, and for it to not take years to add ROCm support. If RDNA5 is the last "RDNA", then there probably is no RDNA6, based on the video.

I take it that's based on the old MLID AT0 XL leak? 154 compute units, 384-bit, 12x 3GB GDDR7. It's presumed to match a hypothetical RTX 6090, but we don't know a lot about that, including the VRAM. RTX 5090 is 512-bit, 16x 2GB modules. If they keep the bus width, there's nothing stopping an RTX 6090 from shipping with 48 GB GDDR7. https://www.pcgamesn.com/amd/radeon-rx-10090-xt-guide https://hardwaretimes.com/forget-the-rx-9900-xtx-rdna-5-based-10900-xt-may-feature-36-gb-vram-10k-cores-rumor/ AMD could justify making such a card if they are selling the silicon for workstation/AI too, which is what the leak alleged. The memory bus is actually cut down from 512-bit. The leaked document also refers to as-of-yet unreleased 4GB GDDR7 modules, allowing an enterprise/workstation product to reach 128 GB on 512-bit (32 modules in clamshell configuration). Another aspect is that RDNA5 appears to be ditching L3 Infinity Cache in favor of a smaller but faster L2 cache, like Nvidia has adopted. Only 40 MiB mentioned for the RTX 6090 competitor (Panther Lake iGPU has up to 16 MiB for graphics). RDNA5 appears to be well on the road to becoming "UDNA":

@taffernicus https://chipsandcheese.com/archive https://chipsandcheese.com/p/hot-chips-2026-samsungs-processing Lot of very in-depth articles at Chips and Cheese, check it out if you dare.

Haven't watched your vid but I recognize Pimax and agree about foveated rendering. I usually see conservative estimates of how much load it would take off the GPU, like 50%, but I think 90% is plausible. IIRC there is a monitor with eye tracking that could use foveated rendering in a non-VR context. It's more complicated because it would need to track your head as well as your eyes, and you would make it less effective if you sat further away from it. I think the case could be made that foveated rendering should take off in displays as well.

Apparently, the crosspoint structure could not be scaled up very well like 3D NAND has been, so it just couldn't keep up with the progress of ever cheaper NAND. Now we have schemes like High Bandwidth Flash that rearrange NAND to give it as much memory bandwidth as possible for AI: https://www.sandisk.com/company/newsroom/blogs/2025/scaling-beyond-the-wall-inside-sandisks-high-bandwidth-flash-for-ai

Not a game, but you get this Witcher 3 concert in up to 4K for free: https://www.gog.com/en/game/video_game_show_the_witcher_3_wild_hunt_concert_2016 https://www.gog.com/giveaway/claim Offer ends in: 1d 14h 04m 01s

I "vibe code" the old/amateur way by asking for specific functions to be made. But from what I hear, the most advanced workflows encompassing entire projects and actually spending money on frontier model tokens are yielding stupendous results. And it will only get better, and probably fast, because coding is a major area of focus, and adjacent to the effort of trying to use models to create better models. On local AI, the RTX PRO 6000 Blackwell 96 GB price has doubled to $16,000, Gorgon Halo 192 GB will probably be at least $5,000, and you can buy a GB300 DGX Station with a 748 GB memory pool for around $95,000. I believe those kinds of capabilities will become dirt cheap after 3D DRAM takes off.

There's a LOT of advanced memory news coming out this week because of Hot Chips 2026. https://wccftech.com/sk-hynix-advanced-packaging-technologies-intel-emib-next-gen-hbm-memory/ https://wccftech.com/samsung-leverages-advanced-logic-processes-to-advance-hbm-works-on-3d-stacked-zhbm/ https://wccftech.com/micron-says-ai-memory-wall-worsening-compute-outruns-hbm-bandwidth-by-3x-every-two-years/ https://wccftech.com/d-matrix-raptor-3d-dram-achieves-sram-class-bandwidth-at-1-10th-the-hbm-power/

I've heard about PIM and HBM-PIM for years. It's vaporware AFAIK, or has a limited scope, otherwise we'd be seeing orders of magnitude improvement and lots of PIM headlines. 2023 article: https://semiconductor.samsung.com/news-events/tech-blog/hbm-pim-cutting-edge-memory-technology-to-accelerate-next-generation-ai/

5050 -30% is what I guessed based on pre-release info. Hardware Unboxed's recent testing shows the RX 9050 8GB at 83% of the 5050 at 1080p, and then it was equal at 1440p. I don't know what could cause that, maybe PCIe 5.0 x16 vs x8, or 32 MiB L3 vs. 24 MiB L2 cache? Not that it's a good 1440p card, but something is going on there. The 9050 8 GB is a real stinker, but I think people would start buying it if it fell under $250. Maybe higher if GDDR6 pricing deteriorates.

People just aren't that desperate yet to drive up ancient, low capacity memory for retro computing. Used DDR4 isn't even that expensive IMO (in the US market, at least). Realistically, I don't think most people want to go older than Haswell (which gets AVX2) with DDR3.

Have immersive sims gone too far?

AMD doesn't treat 6-core/budget X3D very seriously. Their strategy has been to release the single-CCD flagship first, and then release cheaper versions much later. For example, the 7800X3D launched on April 6, 2023, but the 7700X3D launched exactly one month ago, on July 16, 2026. Even when they finally launched 6-core X3D parts, they were often exclusive to specific retailers or regions. Such as the 5600X3D and 7600X3D at Micro Center in the USA, and 5500X3D as a Latin American exclusive. Intel is poised to compete directly with X3D using bLLC-equipped Nova Lake CPUs. The injection of some competition into the market could encourage AMD to compete harder in response, and release budget X3D products with greater frequency and volume. It's hard to give good advice right now because you have to make a judgment call on whether or not the AI bubble will pop, wars will cause supply chain disruptions, and where DDR5 pricing is headed. But if you are considering AM5 and wait for the Nova Lake vs. Zen 6 faceoff, things could get more interesting. Zen 6 will move to a 12-core CCD, so budget CPUs could end up having 8 or 10 cores (a 6-core is still rumored, but I think that could be low volume or OEM-only).

AMD has a nice $350 Anniversary Edition with your name on it.

Zen 6 X3D and Zen 7 X3D could be insane, I think. Since they are actually increasing L3 cache for the first time since 5800X3D. On the same AM5 socket, even. Aluminum's experience shows the importance of 1% lows. So are you hooking this up to a 4K 120 Hz TV and a 360 Hz monitor, or what?

I believe the consensus is mid-late 2027 for AMD RDNA5, and possibly well into 2028 for Nvidia RTX 60 series. It likely depends on pricing and availability of ~36 Gbps 3 GB GDDR7 modules, and Nvidia's willingness to give a crap about the gaming segment. Nvidia's 50 Supers have been leaked for a long time and could act as a stopgap half generation, but they are completely stymied by 3 GB GDDR7 pricing. BTW, Nvidia just raised the price of the Blackwell Pro 96 GB to $16,000: https://www.techpowerup.com/351549/nvidia-rtx-pro-6000-blackwell-96-gb-gpu-now-costs-usd-16-000 It's presumed that 3 GB costs $60-70. So an RTX 5070 18 GB could cost $420 (if not more) for just the memory, and that is not counting anything else or fat margins on top. This advancement (TurboQuant) was overhyped and only affected one aspect of the process. https://wccftech.com/googles-turboquant-made-everyone-panic-about-memory-demand/ Even if it did significantly reduce memory requirements, Jevons Paradox would likely rear its ugly head. Many came to similar conclusions, but I don't think the situation will be that bad. The push for consumer 8K has failed or been delayed by years/decades, nobody cares about it. Cards are often powerful enough for decent 1440p/4K gaming at reasonable settings, with mandatory "better than native" AI upscaling (DLSS 3/4 or FSR 4). Many people are still fine with 1080p at 60 Hz. Consider that 4K120 is 8x the pixels per second. Weaker CPUs are also fine for lower frame rates. We are likely to see a handheld version of the PS6 next generation pushing as low as 1080p30 with relatively weak hardware. If you care about refresh rates significantly above 240 Hz, frame generation will be the feasible way to hit them. Some of the generational stagnation is down to waiting/settling on nodes. Nvidia 40 and 50 series consumer GPUs use the same node. RTX 5090 brute forces its performance by using twice the die area of the 5080, and a significant increase over the 4090, putting it close to the reticle limit. Chiplets can potentially be a way to break the reticle limit while optimizing costs for all segments. RDNA3 used chiplets, RDNA4 switched back to monolithic. RDNA5 is expected to have chiplets shared between dGPUs, APUs, and professional products. You won't see multiple GCDs soon, but the movement is in the right direction. I don't think your "complexity" argument actually makes sense. And it could be disproven by GPU technical details like RDNA5 moving from Infinity Cache (typically 32-128 MiB L3) to smaller, optimized L2 caches as low as 10 MiB. Separate from the AI memory apocalypse, we're seeing a slowdown of Moore's law, more expensive process nodes even before demand is considered, and lackluster competition. There's still nodes planned through around 2041, including new types of transistors. There's a 3D revolution waiting to happen in both memory and logic. Intel may exit gaming dGPUs, but could enter into the area of mega APUs with Razor Lake-AX. Chinese companies are also making viable gaming dGPUs. Gaming requirements are generally aligned with the console market. If you can match or beat the estimated performance of the PS6 handheld (which may include some advanced features such as neural texture compression), you are at least good for 1080p gaming throughout the next 10-15 years. Because games will continue to target PS5 for a few years, and then target the PS6 for a few years even if a PS7 arrives in the late 2030s.

It's a Thief-like game, not AssCreed.

They are supposedly going to be using chiplets again, although only 1 GCD in anything consumer-related AFAIK. The big leak several months ago is that some mobile APUs could share a chiplet with low-end discrete GPUs, using LPDDR5X in both cases. Also, Xbox Helix should be sharing a chiplet with mid/high discrete.

@revelator The really sad part is that if it dropped to around $220 or less, it might be good price/performance in the US market. It looks like an RX 6600 with newer features and better raytracing. But of course, it's not available in the US or Europe right now. The 4 GB card isn't relevant right now. You could see it as an OEM's hedge against extreme memory pricing forcing even a GDDR6 death spiral. AMD went with the worst of all worlds by not making it a 75W card, just slightly above that at 92W, and wrecking the die. Just a measly 24 compute units would have made a much more compelling product. Oh well, on to RDNA5. Here's some hype for that:

Btw the National Recording Registry (Library of Congress) inducted Bobby Prince's Doom soundtrack this year. https://www.engadget.com/2173357/the-original-doom-soundtrack-is-officially-in-the-library-of-congress/ https://newsroom.loc.gov/news/national-recording-registry-inducts-sounds-of-taylor-swift--beyonc---the-go-go-s--vince-gill--weezer/s/bc258688-e655-4ffb-9f91-f32b94956f36

I'll go for it.

https://videocardz.com/newz/amd-radeon-rx-9050-launches-at-279-in-asia-pacific-japan-and-latin-america https://videocardz.com/newz/amd-confirms-radeon-rx-9050-4gb-exists-but-it-is-oem-only https://videocardz.com/newz/sapphire-launches-radeon-rx-9050-pulse-with-8gb-memory-and-100w-tbp RX 9050 8 GB: 16 CUs, $279, limited to Asia-Pacific, Japan, Latin America. AsRock and Sapphire models spotted. RX 9050 4 GB: 16 CUs, OEM-only, in a single pre-built PC. They are disabling 50% of the Navi 44 die to make these, so it is a waste of silicon by any stretch of the imagination.

Got another email from GOG, linking to this video on their secondary channel:

Here's one way in which they have something to do with each other: if the remaster is considered a stinkbomb, GOG is unlikely to pull the original version off the "shelf". Which has happened in the case of some remasters IIRC. Obviously, the community is obsessed and the games are well preserved. I see this as another aspect of that. I may have even voted for Thief Gold (don't remember). I think GOG has attempted to make bug fixes to games in the Preservation Program, but probably not to a great extent whatsoever if they can only afford a couple of programmers and often don't have the source code. In this case they will just defer to NewDark. I think not having to install NewDark is a potential benefit for GOG users, but it was probably already the case, idk.