Wednesday, March 11th 2026
AMD Prepares "FSR Diamond" Update for Xbox Project Helix
AMD is reportedly developing a next-generation FSR update, codenamed "FSR Diamond," for the upcoming Xbox project "Helix" gaming console. With the Project Helix console expected to launch in 2027, the details of "FSR Diamond" remain a mystery. It's unclear what AMD aims to achieve with this AI-powered video generation technology, but it likely builds on previous advancements like Radiance Caching and Ray Generation, adding a new dimension to the company's graphics capabilities. We might see an AMD equivalent of the multi-frame generation technology found in NVIDIA's GeForce RTX 50-series GPUs and Intel's XeSS 3.0. Since Project Helix is anticipated to feature RDNA 5 / UDNA graphics IP, this feature could be exclusive to that generation, as AMD tends to tie its latest technologies to the current RDNA IP.
AMD already differentiates its latest FSR "Redstone" suite of technologies, with features like Ray Regeneration and Radiance Caching exclusive to RDNA 4 hardware in the Radeon RX 9000 series of GPUs. Other basic technologies, such as upscaling and frame generation, are supported on older RDNA 3/2/1 generations but use an FSR 3.1 fallback, with no FSR 4 support currently available. However, since INT8-based FSR 4 exists, it may only be a matter of time before the company extends this capability to older GPUs, though the expected performance might not be optimal. For multi-frame generation and potentially dynamic multi-frame generation, "FSR Diamond" would need specialized hardware. Even NVIDIA, with its MFG 6x mode and Dynamic MFG, keeps those features exclusive to the GeForce RTX 50-Series "Blackwell," which uses hardware flip-metering available only on the newest GPU generation. Similarly, RDNA 5 / UDNA could incorporate these hardware components as well.
As we anticipate features like next-gen neural rendering, ML-based upscaling, multi-frame generation, and advanced ray tracing for the upcoming Xbox, known as Project Helix, not much has been revealed about the console itself beyond its logo and project name. There have been longstanding rumors about the next-gen Xbox, with recent speculation suggesting it will be a hybrid device that bridges the gap between a living room console and a PC.
Sources:
Jack Huynh on X, via VideoCardz
AMD already differentiates its latest FSR "Redstone" suite of technologies, with features like Ray Regeneration and Radiance Caching exclusive to RDNA 4 hardware in the Radeon RX 9000 series of GPUs. Other basic technologies, such as upscaling and frame generation, are supported on older RDNA 3/2/1 generations but use an FSR 3.1 fallback, with no FSR 4 support currently available. However, since INT8-based FSR 4 exists, it may only be a matter of time before the company extends this capability to older GPUs, though the expected performance might not be optimal. For multi-frame generation and potentially dynamic multi-frame generation, "FSR Diamond" would need specialized hardware. Even NVIDIA, with its MFG 6x mode and Dynamic MFG, keeps those features exclusive to the GeForce RTX 50-Series "Blackwell," which uses hardware flip-metering available only on the newest GPU generation. Similarly, RDNA 5 / UDNA could incorporate these hardware components as well.
45 Comments on AMD Prepares "FSR Diamond" Update for Xbox Project Helix
I will *love* to hear AMD's explanation of whatever it is that won't be supported prior to RDNA5 if that happens to be case
Apples and oranges. Get back to me when DLSS5 is entirely restricted to the RTX 5000 series and you'll be on equal footing.
Pretty sure FSR3+ is supported to everything including competeing brands, only the machine learning FSR4 is not.
Obviously it should but the statement "gatekeeps EVERYTHING" is pretty silly.
5090owners can shut up about it.That’s far too grown-up a conversation to have with users who know nothing better than to show up in these threads to bicker and derail the discussion.DLSS 4.5? Available on 8-year-old RTX 20 series.
FSR 4? Artificially restricted to RX 9000 series.
Mega Geometry? Available on 8-year-old RTX 20 series.
AMD's equivalent cluster-based ray tracing solution? Nowhere to be seen
Studio drivers? Available on 8-year-old RTX 20 series.
AMD? Discontinued driver development for RX 5000 and 6000 series, placing them in "maintenance mode" while only issuing current branch drivers to the last two GPU architectures
Oh, and... if Kepler's right on the money, remains to be seen. But if this is any precedent to be set? RX 9000 owners shouldn't expect any diamonds
I could go on and on, and on, but it just remains: Nvidia is gatekeeping pretty much nothing that previous generation hardware can support. The only things that haven't been backported to Ada and earlier are features that require Blackwell's hardware-based flip metering capabilities, I believe that is Reflex 2 with Frame Warp (which is still under development) and dynamic rate frame generation.I love you too, but why the strikethrough? I do have an RTX 5090.
Don't ask questions, just consume product and then get excited for next product.
When all the user base actually wants is INT8 FSR4 for their RDNA2-3.5 cards, and better official FSR4 support and [more] implementations generally.
AMD is, CURRENTLY, gatekeeping everything. Meaning they have gone backwards. But we're not willing to call AMD out because they're the poor $35 billion underdogs.It's neither. It's greed. We all know that INT8 for FSR4 exists, and AMD chose not to support it. This decision here only confirms their course of action: you only get support so long as you own the current GPU, once the new gen is out you no longer matter.
The very thing AMD fanboys accuse nVidia of doing, is now AMD's business model. It was never about FP8 hardware, it was always about forced obsolescence.
ReBAR? Not available on 20 series because reasons. AMD supports it on older series because it's a standard PCIe feature.
FG? Not available on 20 and 30 series because reasons. AMD FSR FG runs on those cards proving those cards can in fact run it even without dedicated hardware support.
MFG? Not available on 20, 30 and 40 series despite 3rd party programs proving those cards can in fact run it even without dedicated hardware support.
Adaptive MFG? Available only more than a year after launch on 50 series exclusively.
Neural Rendering? Likely exclusive to 60 series.When AMD supported everything, including Nvidia cards with FSR 3.1 then the argument was that a solution targeting all cards does not look very good and that AMD needs exclusive features to differentiate itself and entice people to buy their hardware. Now that they do they're the bad guys for doing what people wanted them to do?
The hypocrisy is strong here...
Regarding DLSS upscaling, you do realize that this is far more than AMD has done for their customers, right?
Resizable BAR is not an essential feature, nor is it globally enabled, the aperture size being too large reduces performance in games. Graphics drivers profile titles that show performance regressions (which is most of them) and disable the feature accordingly. Stands to reason adding ReBAR support to Turing wasn't necessary, assuming you're running a Titan RTX you're probably leaving 0-3% of performance on the table.
The frame generation things, requires Blackwell flip metering. It's a hardware feature. Another thing that requires Ada or newer is shader execution reordering, which, again, is a hardware feature
Neural rendering is already available on all RTX GPUs, 20 series and up, running driver 572.16 or newer. It's been around for about a year now
The one thing I'll give you is that some sort of FG could be added to earlier generation cards, yes. However, given how bad software-based frame generation tends to be (just look at FSR 3 or LSFG ghosting), I'm not holding it against anyone. Besides, if I want to use FG that badly, I already own LSFG on Steam
You know, the new RDNA 2 compatible upscaler that Sony says is FSR4.
How come Sony can do it and AMD can’t?
I thought multi frame generation for RDNA 4 was already hinted at with whatever was found in the recent AFMF?! It was mentioned in an interview that AMD would port over whatever made sense, and the recent Microsoft statement about "Build for Xbox on PC" is interesting, alongside Sony's decision to pull back from porting over to PC. Hopefully they'll be overlap going forward (outside of Optiscaler)...
www.theverge.com/news/893196/microsoft-says-you-should-build-next-gen-xbox-games-by-building-them-for-pc
www.techpowerup.com/346886/insider-confirms-sony-ditching-pc-ports-youll-be-seeing-fewer-single-player-games-arrive-on-pc?cp=1
"deeply integrated into the SDK?!"
insighttechdaily.com/amd-fsr-4-1-dll-leak-game-integrated-ai-upscaling/
www.wheresyoured.at/the-beginning-of-history/
Multi FG (and now Adaptive FG) uses Tensor cores. Tensor core are there since 20xx.
Adaptive FG works with Lossless Scaling utility on a very wide range of GPUs.
Now tell me about gatekeeping.DLSS is a name used to a bunch of things.
Some of them ARE restricted to rtx 50xx only.
You do know how processor architectures work, right?See this tells me you are not reading what I write and are just arguing for the hell of it.Well you see, that doesnt exist, just like INT8 mode doesnt exist, because that undermines their argument that AMD can do no wrong and definitely isnt a greedy corpo trying to force them onto newer hardware by artificially gatekeeping features. So that will be ignored or downplayed.
But its nVidia that has crazy mindshare that defends them no matter what.Are you talking about yourself there?
Yes, FSR 3 looked bad. It compared horribly to DLSS3. It also compared badly to XeSS.
Yes, having features exclusive to AMD GPUs can entirce sales.
How TF do you get "AMD is justified in cutting off their OWN USERS with new software" out of that argument? This is a wild leap in logic.
However, yesterday at the GDC, there was a lot of discussion about the actual direction where the next-gen hardwares are heading, like RDNA 5. And the situation is that this direction largely aligns with addressing the criticism that was first raised about neural rendering.
To put it very simply, neural shaders can work with an enormous amount of data, but it is important how fast a GPU is capable to process this data. Every architecture has a maximum data processing window, meaning that as long as a neural shader workflow stays within X amount of data, the architecture can approach the theoretical maximum of its computational performance. If we take the widely used FP8 format, all of the current GPUs (even the RDNA 4 and Blackwell series) has a data window well below 100 kB for what a neural shader can process in one workload. This limits what can these methods achieve, because they can only process a small tile of the full frame at a time, which leads to significant losses during computation. In other words, until this data window is increased drastically, the achievable graphical quality will remain very limited with neural shaders.
RDNA 5 is the first architecture that actually responds to this problem. We did not receive official figures at GDC, but it was mentioned that the limit, that is currently well below 100 kB, will increase to several megabytes. It will likely be somewhere around 25 MB, and the reason there is no exact figure yet is that it is reportedly configurable, so it can be configured to the specific workloads. In neural rendering, this would make RDNA 5 incomparably better than any current GPU. If a neural shader is written for to works with for example 10 MB of data tile, the RDNA 5 can still utilize roughly 90% of the available computational performance. If you run a neural shader that works with 10 MB data tile on an RTX 50 series GPU, the hardware will be able to use only about 2% of its theoretical computational performance, because during the remaining 98% of the workload the matrix units are simply waiting for data.
So yes, we can say that current hardware will also support neural rendering, and on paper that is true. But under any neural shader that works with larger data tiles, they will be 100,000 times slower than next-gen GPUs, such as those based on RDNA 5.So Sony has an advantage on this one, because they modified the RDNA 2 design to combat a very specific problem of these ML workloads. The original RDNA 2 can't use those intrinsics ops that made PSSR 2 possible on a PS5 Pro. And this is why Sony didn't release PSSR 2 on PS5, beucause it is impossible to execute efficiently even the modified FSR 4 code.
The problem with FSR 4 is the heavy ML workload, which is designed to RDNA 4. The Int8 variant use smaller tiles, but if you use smaller tiles, you will need to calculate much bigger tile overheads, so most of the computations will just not used for the actual results. Where I work, we are also developing our own upscaler for cooperative vector, and we are probably struggling with the same problem as AMD. If the ML workload is heavy, the tile size has to be determined very precisely, otherwise we lose a lot of performance due to memory accesses.
RDNA 4 allows for larger tile sizes thanks to its improved memory subsystem compared to RDNA 2 and 3. However, if we try to tune the tile size for RDNA 2 and 3, then about 60–70% of the work becomes unusable, due to the mandatory tile overhead.
After this we came up an idea to make the ML workload itself significantly lighter on RDNA 2 and 3, which allowed us to reach the tile size used on RDNA 4. However, this created another problem. Even though the upscaler is technically ML-based, its quality does not even reach the level of FSR 3. For this reason, we also ended up dropping RDNA 2 and 3 entirely, because the hardware design is simply too limited for this purpose. We also asked Sony for help, since they have a lot of experience with PSSR on RDNA 2, and they told us to drop it, because the hardware is only suitable for this with the modifications they made for the PS5 Pro.
DLSS is vendor locked just as FSR4 is. Also like i have said dozens of times in TPU forums before about Nvidia offering DLSS to older series (at worse performance): Should i just forget all their prior transgressions because they've finally done one thing right?
Should AMD also be forgotten if they suddenly decide to offer FSR4 INT8 on RDNA2 and RDNA3?
No. No they shouldn't. But i suppose things have gotten so bad for consumer hardware that people will celebrate even small victories whenever they can get them.Ah yes the classic Nvidia apology: "not an essential feature" and "small performance uplift". ReBAR is a standard PCIe feature that can offer as much as 10% performance uplift. AMD enabled it on their older series too while Nvidia is (still?) using a whitelist method.
I guess we finally found one aspect where AMD's hardware is better. I would think that these days ReBAR ON should be the default.
Besides adding it to 20 series is easy for Nvidia.Interesting. People harp on AMD for not supporting FSR4 on older series because those lack appropriate hardware to run it at the same performance and image quality as RDNA4, but it's ok for MFG to be 50 series exclusive because it's a hardware feature?
Either criticize both of them equally or criticize neither of them.But, but it's "not an essential feature" and "small performance uplift" no?I didn't know that. Thanks for the correction.I used FSR FG on my 20 series card when i still used it prior to 2025. If it had ghosting i wouldn't have used it.But it's not a native solution. Isn't that the criticism directed towards AMD users relying on OptiScaler for example?
If it's ok for older RTX series owners to use FSR FG or LSFG, then i suppose it's ok for AMD users to use OptiScaler to get FSR4 support both in older games, Vulkan games, DLSS games and on RDNA3.You mean works with compromises. That's the tradeoff DLSS 4.5 is also making on older RTX series.Replace the words INT8 and AMD with FG and Nvidia and it's the same thing.
No one is saying AMD is not not a greedy corpo gatekeeping, because that's what they all do.
AMD already tried the open source "supports all GPU's" approach. How did that work out? Were the forums filled with praise or were people (primarily Nvidia users) still bitching and moaning about image quality?
Now AMD is trying the closed/exclusive route with much better image quality and people are still bitching and moaning, because now it's locked.
You cant have both. You can have broad support with worse performance/image quality or targeted narrow support with good performance/image quality.
Same thing with prices. When AMD was half as expensive while being at performance parity people still bought Nvidia and AMD almost went bankrupt.
AMD got the message loud and clear and now offers -50 vs Nvidia, because people still buy Nvidia., but at least now they're in no danger of going bankrupt.Nvidia has cut off their own users for generations. Maybe the criticism finally paid off that they decided to make a compromised DLSS 4.5 for older series.
As users we can only hope that criticism towards AMD can result in a similar outcome somewhere down the line.
I just cant stand the hypocrisy of criticism one for which the other one gets a pass.
I couldn't care less about the past. I care about the here and now, as well as the future. The past at best can be used to infer a behavioral pattern for the future, yet paradoxically all of AMD's good past behavior has not been repeated with the current generation of products
Interestingly, the complete 180 on the good guy AMD act actually accelerated their market share loss, no less than what they deserve.