Monday, May 5th 2025
AMD Patents Provide Early UDNA Insights - "Blackwell-esque" Ray Tracing Performance Could be Achievable
Last September, AMD leadership publicly revealed UDNA—an "unforking" of previously separate enterprise and commercial GPU branches. Not long after this announcement, TechPowerUp's resident Serbian correspondent—AleksandarK—sat down with Team Red's Andrej Zdravkovic. The Chief Software Officer (and SVP) stated that a fair chunk of UDNA-related development work would be done by local engineers. Zdravkovic discussed this technology's eventual deployment in futuristic "AI PCs," but gamers have been salivating at the prospect of a proper successor to RDNA 4. A next-gen graphics architecture seeker—MrMPFR—has combed through official documents for any sign of UDNA preview material. The noted /Hardware subreddit member managed to distill their initial (very long) set of findings into an "easily digestible overview." They stated that this was just a small case of: "reporting and a little analysis on AMD's publicly available US patents filings," and other public-facing resources/archives.
Gleaned information included: "finalized architectural characteristics in future RDNA generations, AMD DXR IHV stacks (driver agnostic), and AMD sponsored titles. But please take everything with a grain of salt given my lack of professional expertise and experience with Real-time ray tracing (RTRT)". MrMPFR believes that Team Red started picking up former NVIDIA and Intel engineering talent, back in 2022/2023. In addition, a lot of new hires were apparently sourced from academic institutions. In theory, these newer team members have not had the time to make major inroads—in terms of getting finalized products out into the wild. MrMPFR reckons that noticeable contributions will accelerate AMD's making of "RDNA 6+/UDNA 2+," and beyond. Early 2025 leaks have pointed to the company collaborating with Sony; their "PlayStation 6" console is tipped to be powered by some fork of Team Red's "UDNA" graphics technology.
MrMPFR unearthed a wealth of intended ray tracing performance improvements—a key goal being the superior usage of BVH (Bounding Volume Hierarchy) management. The keen industry observer shared the benefits of this feature: "(BVH) allows for lower CPU overhead and VRAM footprint, higher graphical fidelity, and interactive game worlds with ray traced animated geometry (assets and characters) and destructible environments on a mass scale...the patent filings cover smarter BVH management to reduce the BVH construction overhead and storage size and even increasing performance with many of the filings, likely an attempt to match or possibly even exceed the capabilities of (NVIDIA's) RTX Mega Geometry (tech)."
Their TL;DR part #2 section compared forthcoming developments with a familiar rival current-gen solution: "the patents indicate a strong possibility of almost feature level parity with NVIDIA Blackwell in AMD's future GPU architectures likely as soon as 'RDNA 5/UDNA' based on the filing dates. We might even see RT perf parity with Blackwell at iso-raster perf, that's an identical FPS drop percentagewise between architectures...If more architectural changes make their way into next-gen RDNA than those afforded by the current publicly available patent filings then it is very likely to exceed NVIDIA Blackwell on all fronts, except likely only matching ReSTIR path tracing (PT) and RTX Mega Geometry functionality. If this is true then that would be AMD's 'Maxwell moment' but for RT."
Sources:
/Hardware Subreddit, Wccftech, TweakTown, XDA Developers
Gleaned information included: "finalized architectural characteristics in future RDNA generations, AMD DXR IHV stacks (driver agnostic), and AMD sponsored titles. But please take everything with a grain of salt given my lack of professional expertise and experience with Real-time ray tracing (RTRT)". MrMPFR believes that Team Red started picking up former NVIDIA and Intel engineering talent, back in 2022/2023. In addition, a lot of new hires were apparently sourced from academic institutions. In theory, these newer team members have not had the time to make major inroads—in terms of getting finalized products out into the wild. MrMPFR reckons that noticeable contributions will accelerate AMD's making of "RDNA 6+/UDNA 2+," and beyond. Early 2025 leaks have pointed to the company collaborating with Sony; their "PlayStation 6" console is tipped to be powered by some fork of Team Red's "UDNA" graphics technology.
Their TL;DR part #2 section compared forthcoming developments with a familiar rival current-gen solution: "the patents indicate a strong possibility of almost feature level parity with NVIDIA Blackwell in AMD's future GPU architectures likely as soon as 'RDNA 5/UDNA' based on the filing dates. We might even see RT perf parity with Blackwell at iso-raster perf, that's an identical FPS drop percentagewise between architectures...If more architectural changes make their way into next-gen RDNA than those afforded by the current publicly available patent filings then it is very likely to exceed NVIDIA Blackwell on all fronts, except likely only matching ReSTIR path tracing (PT) and RTX Mega Geometry functionality. If this is true then that would be AMD's 'Maxwell moment' but for RT."
60 Comments on AMD Patents Provide Early UDNA Insights - "Blackwell-esque" Ray Tracing Performance Could be Achievable
- TSMC begins production of 2-nm wafers, with orders open from April 1, 2025.
- Large-scale production is expected by the end of 2025, with plants in Hsinchu and Kaohsiung.
www.turtlesai.com/en/pages-2612/tsmc-starts-2nm-wafer-production-apple-among-firstGiven Navi 48's transistor density of over 150 MTr/mm2, it doesn't make sense to use 3nm. It will be a half-node shrink.
In order to improve the ray-tracing performance, they need to put more RT units, or to design a separate chiplet exclusive for RT only.
Also, don't get why they need so slow and expensive "real-time" RT, when all modern games anyways look like shit.
Edit: It would be better to backport UDNA to 5nm and make it double the size of Navi 48 - around the reticle size limit.
Blackwell has been lambasted for lacklustre gains over prior gen, driver issues, missing rops, melting connectors, low VRAM, bad availability, high power consumption and high prices - it's literally being called the worst NVidia launch ever.
Ada was heavily criticised for pricing, the lineup (resulting in un-launching a card) availability, low VRAM, the melting connector on 4090's
All these sentiments are from tech press and the community alike, it doesn't sound like being hailed as heavens gift to me.
There are some products that attract some praise for their performance level in there, but your statement is a serious dose of hyperbole.
As to AMD being a generation behind where the Nvidia products attracted certain praise in their limelight, I don't see what's surprising there. When they launched they had the best performance in X metric. Now times have changed and the bar has been raised, ergo that level of performance just isn't as impressive as it used to be. AMD did a fantastic job disproportionately increasing RT perf with RDNA4, I hope they can do this again with UDNA as it'll go down extremely well with reviewers and users.
www.amd.com/en/newsroom/press-releases/2025-4-14-amd-achieves-first-tsmc-n2-product-silicon-milesto.html
I made this case before and I'll say it again... for the GPU realm, I believe it's the wrong move, even 2 years from now. These latest generation nodes serve no purpose for home computing anymore. The products are in extremely high demand in stupidly profitable, very high margin, reliable corporate segments. Unless you want the GPU pricing situation to continue and worsen over time, forget about these hyper fancy 3 and 2 nm chips. Make stuff with current 7 and 5 nm technology, it's better for all of us.This rhetoric was weird last year and it's even weirder now. Corporate brand allegiance is something only a dimwit has, GPU vendors are not sports clubs, you don't have to feel proud you have an AMD or Nvidia GPU. This behavior is the primary entry port to justify abusive pricing, especially when there is a large gap in both companies' products. You want to know why I purchased an RTX 5090 over a 9070 XT? Here, this is why:
That's really all there is to it, and at that scale, it was worth both the price (which was my regional MSRP) and even the excruciatingly long wait for it. AMD addressed most of my technological concerns with the 9070 XT, such as the media encoder and OpenGL performance. What they are sorely lacking right now is a compute runtime that is as seamless and broadly supported as CUDA. They are aware of it, they are working on it, they will get there. It takes time. What they cannot offer me right now is the performance I want. The 9070 XT isn't better than the card I had, and the vanilla 9070 barely matches the one I had 5 years ago. Not exactly appetizing, yet... these performance figures are actually excellent, and there is no game that it won't masterfully run right now, at very good resolutions and frame rates. The largest majority of people should be very happy with it, and from the looks of it, they are, because all reports on the internet on both specialized and general media point out that these cards are selling exceptionally well and AMD is recovering market share, as expected.
Once they up their driver release schedule, get ROCm or whatever bundled runtime going, and increase the performance to match NV's, then I'll happily buy another AMD card for my main machine. I need to add though, you missed the critical issue: it's not that gamers pay for expensive cards. We're irrelevant, even as a collective group. They just have next to no incentive to build and maintain gaming cards when they can get the same silicon and sell it to the AI customers at triple the price and half the effort.
GPU vendors are like political parties - one is better for you, another is better for me. Nvidia is greedy and extremely anti-people, while AMD tends to support the people's needs more.
Look at how there are much less AMD Radeon graphics cards which run out of VRAM, exactly the opposite of Nvidia's history of almost always lying to the people and giving less for much more.
Now, i did a little digging after i heard some youtubers, even digital foundry, how AMD does raytracing denoising in SOFTWARE while Nvidia has some specialized hardware just for that, that is true for RDNA2 and 3, RDNA4 has some hardware denoising capability that right now IS NOT USED IN ANY GAME.
Once AMD gets it's hardware raytracing denoising in some games then things should look very well, but, you will never see AMD hardware raytracing denoising in cyberpunk 2077, black myth wukong and some other titles sponsored by Nvidia, just like you won't ever see FSR4 in these titles.
Another reason Nvidia does so well in raster with less VRAM and less power is because they have some advantage in texture compression, as i did a little digging the most used algorithm is BC7 in which Nvidia is better and offers a little more than AMD, outside standard spec, BC7 is used for compatibility with older GPU's.
These technologies are obsolete but are still used for now, both Blackwell and RDNA 4 get ready for neural texture compression, neural raytracing denoising and whatever they agreed for DX13 and PS6.
As usual Nvidia won't want standards because they want to charge the most and be top dog, so they will agree to something standard with Microsoft and AMD and then add something unique to them, then spend a lot of money with game developers and promoting that unique feature.
And that is how you get a monopoly, be unique, be different.
Nope, the 5070 Ti is 21% faster in ray-tracing, which also makes it similarly more power efficient (304W vs 300W). ~20% aligns with my other post. 20% difference is not matched (~5% I may call matched). Also, the 5070 Ti (or any Ada/Blackwell GPU (Blackwell is just an Ada refresh)) is ~50% faster and similarly more power efficient (see Source 3) (I don't see path-tracing benchmarks by TechPowerUp for the recent GPUs, so I have to reference other sources) in the visually stunning path-tracing, than the 9070 XT. Compared to a RDNA3 GPU, the difference is 155% or 2.55 times (again, same Source 3: (355W [7900 XTX]/10.2FPS)/(300W [5070 Ti]/22.0FPS)), so AMD has improved nicely and was lucky that NV didn't improve with Blackwell.So AMD is trying to catch up on the missing 20%, but the Blackwell successor will be faster again by like 20-33% and AMD would be behind one generation again? Ok.
So future game reviews and hardware reviews should spell it out for everyone, the bullshit about architecture doesn't stand anymore, it's all about who is better at anticompetitive practices.
The reason is that in the past it was not really necessary to access VRAM as one large pool and 256MB chunks were fine.
So calling it trial is odd. Every little but helps and it's free performance by enabling an old PCIe feature that was not used before.
My point was that Nvidia shafted it's customers and refused to provide a simple feature that itä's competitor enabled for much older cards.That's verifiably false. If 20 series users can run AMD's framegen with good performance and image quality then there's no doubt Nvidia could do much better job if it wished. Their OFA excuse is pathetic. Either AMD's framegen on Nvidia cards entirely bypasses OFA or Nvidia is so incompetent gen after gen that they cant backport features to their older cards because of their constantly underpowered hardware in each gen. Neither looks good for them.RDNA 4 and likely earlier generations will have multi framegen. AMD already proved that framegen does not require specialized hardware. Yes it does benefit from hardware accelerators, but it's not essential. Upscaling is a different story and here it seems hardware acceleration is essential for the best image quality.No they didn't. RDNA 3 was mix of 5nm and 6nm. Not to mention first chiplet GPU and it's associated growing pains. 40 series was 5nm. Thus in my eyes Nvidia had a node advantage. This gen they are both on the same node. Nvidia is better here no doubt but not by miles. For example 9070 XT is a 304W card. 5070 Ti is a 300W card. Both use 16GB of VRAM and both are roughly the same size of chip (Nvidia's is a little larger). 5070 Ti is on average 9% faster while using a slightly larger chip and a bit less power (TPU's GPU database).
www.techpowerup.com/review/asus-radeon-rx-9070-xt-tuf-oc/35.html
But there are also other drawbacks for RTX 5090:
1. Melting power connector;
2. Too large and heavy - sagging issues;
3. 32 GB is an overkill;
4. Extremely overpriced;
5. There are better options on the market;
6. Depending on the settings, there are cases where there is no perceivable user experience difference between the RX 9070 XT and RTX 5090:
www.techpowerup.com/review/asus-geforce-rtx-5090-astral-liquid-oc/11.html
This is a measly 32% performance drop from RTX 5090 to RX 9070 XT.
RT is still almost a no-go on anything less than a 5070Ti, unless 30fps is your bag.
To get FSR 4 in games at this point in time, you're relying on a driver level override onto an FSR 3 entrypoint, it was the only way AMD found to have any meaningful amount of supported titles without actually working with developers to get it done. DLSS is thoroughly documented, has a proper API, allows the user to configure its automatic exposure, training model preset, resolution, etc. - even though this is not always exposed with intuitive controls to the user as they are intended to be developer-facing. But it's there.
Things have never been better at the red camp than they are right now. What you need to do is keep in mind that AMD is not a charity, they do not "genuinely care about gamers", and they are coming from a severely weakened market position, so they have to look appealing to claw back market share. If the tables were turned they would do the same thing Nvidia does.True, yet very much a developer concern. CUDA offers compute capability levels which tier the hardware and makes it very clear to the developer what the hardware they are developing against can support, for example, the RTX 3050 laptop GPU (GA107) supports compute capability level 8.6 - regardless of the CUDA runtime version, it will run everything that level 8.6 can provide and whatever is retroactively added to this compute capability tier. It's the same thing that DirectX does with its feature levels, which is why it's possible to run DirectX 12 Ultimate (12_2) games on hardware like the RX 5700 XT, which doesn't support the extended features of Ultimate, or why you can run a DirectX 12 game on a GTX 600/Kepler GPU with feature level 11_0, or DirectX 11 games on hardware like the Radeon HD 4000 series with 10_1 level.Stop embarrassing yourself.
I really hope they try to be better this time, RDNA4 has it all except software support.
#7. Buggy drivers, it seems like Nvidia is using AI to QA/QC their drivers, and I couldn't imagine paying the green team tax to experience black screens and games crashing.
I think it's a disappointing double standard how no one seems to criticize Nvidia when their drivers are causing issues.
But you can't reason or have your own opinion with users who will always buy Nvidia anyway.
No, they are obviously refactoring driver code. It's the exact same reason AMD had dogshit drivers for the past 4 years and suddenly, it's actually rather good - they've been extensively rewriting their kernel and user mode drivers - something they just got finished with recently. I'll apply to equal measure for both vendors, I know why it's happening, I just fail to see where that is the customer's problem. At least Nvidia is releasing fixes at a very fast pace to make up for it. I know they will make up for it, as 566.36, the last Release 565 driver, literally had no open issues at the time of its release and both Release 550 and Release 535 drivers which were kept around in the enterprise models are extremely polished by now.
They have already announced that the 900 and 10 series GPUs will be retired with CUDA 13.0 and the next driver branch, so I'm positive they are laying a foundation for future driver releases and a minimum baseline for software to target once the newer drivers ultimately drop these products.Positively sure. More than half of the 9070 XT's media presentation at launch was literally them showing that they fixed many shortcomings of RDNA 3 that made those cards markedly worse than their Ada Lovelace competitors... but Blackwell isn't a mere re-release of Ada. There are many improvements to it.
Upscaling, later and worse. Frame gen, later and worse. RT, later and worse. Compute (AI or standard), later and worse, etc.