AMD RDNA4 Architecture Speculation
- Thread starter Frenetic Pony
- Start date
FSR isn’t even vendor exclusive. What would make you think that it would be rdna 4 exclusive?
There’s no common api for running ML models on a GPU?
Maybe I'm mistaken, but I think that (for end-user AI upscaling) you don't need to perform machine learning on your GPU, just inference.
DirectML can run on any GPU, it technically operates in the same space as CUDA, ROCm, and OpenCL.
Specific to this case though, DirectSR should be sufficient, AMD would have to make their solution hardware agnostic. I bit like XeSS DP4A variant runs on any GPU, but XMX variant runs only on Intel Arc cards.
As far as the understanding of rumour mill goes, RDNA4 does not contain Matrix Cores as per their MI line, that won't come until UDNA. So we should be expecting something closer to the solution within PS5 Pro. that could technically run on everything, although I expect RDNA4 cards to run it best.
DirectX-Specs/DirectSR/DirectSR.md at master · microsoft/DirectX-Specs
Engineering specs for DirectX features. Contribute to microsoft/DirectX-Specs development by creating an account on GitHub.
github.com
Last edited:
Sega_Model_4
Newcomer
When they showed Neural Supersampling and Denoising for Real-time Path Tracing, they suggested that it would work on any GPU, but would have better performance on RDNA 3 and onwards.
Yes that’s what meant by “running”.
DirectML can run on any GPU, specific to this case though, DirectSR would be sufficient.
DirectX-Specs/DirectSR/DirectSR.md at master · microsoft/DirectX-Specs
Engineering specs for DirectX features. Contribute to microsoft/DirectX-Specs development by creating an account on GitHub.
How will DirectSR help with enabling older AMD cards or Intel/Nvidia to run AMD’s ML model? At minimum you would need something like DirectML and AMD would have to ship their model with every game and Intel/Nvidia would need to update their drivers to support it. AMD’s older stuff still won’t be able to run it though.
Driver updates need to coincide with DirectML and not the model that AMD is providing. But, yea, legacy games, would all need patching for it to work, the models don't need to be shipped with driver updates. I run nvidia, and we have access to XeSS without driver updates.
I'm just speaking to the possibility that AMD could have gone multi-vendor - as in the option exists, it won't be clear until Jan 6, the more complex the model, the more inference power required. That intrinsically kills off older cards.
The only reason I'm brought this up was because of earlier videos indicating that it would be available on their handhelds. And that is, well, ranging between RDNA2 through RDNA3 so far. So when I ear it's not coming to anything but RDNA4, that's unfortunate as it also kills off console access.
If it's RDNA4 only, I expect this to be a beefier model.
Last edited:
Not going to lie, if they make it RDNA4 only it's pretty much dead. They have 10% market share, or something like that, total. RDNA4 market share will be like less than 1% for a while. What's the incentive to support it with the exception of partnered titles? Seems like a losing strategy. Fully expect it to support a wide range of cards and maybe have RDNA4 optimizations.
Isn’t that version of XeSS running as a regular DirectX shader?
Yeah even if it’s technically possible to ship a DirectML implementation they couldn’t possibly market it as cross vendor. How would they test Intel/Nvidia compatibility?
If RDNA 3 can run the model with acceptable performance I don’t see why they would lock it to RDNA 4.
Intel has 0% market share and XeSS is doing very well.
I also agree with your desire here. If we use 5Pro as a baseline, there's not really a lot of options here if the model requires that much power.
If they are releasing a light weight version of the model for older hardware, like Intel does with XeSS that may work.
DirectSR provides the API to support new methods without any kind of active involvement of the developer.
If we assume this line of thinking, If they built FSR4 on DirectML, it should technically run on anything that supports DirectML. All hardware is capable of running ONNX models, and that's what DirectML supports. I would agree that the model would have to ship with the title however, and it couldn't ship with drivers as they do today.
There's some benchmarking here where DML and Cuda and RocM all seem to trade blows.
DML extends off DX12, if you wanted to get into the SuperResolution game while being vendor agnostic, you would go DirectML.
Last edited:
Pretty much what the rumours have been saying for a while. Add to that 16 GB VRAM and slightly lower power than the 7900XT. Expected price is $599
Price isn't the only issue. As the RX 6600 series showed, despite lower performance, consumers still bought Nvidia over AMD. They might do better than RDNA3, but still don't see it outselling the 5070, even though it will have only 12 GB VRAM.
Agreed, FSR4 would likely run faster on RDNA4. Supposedly it uses FP8 on RDNA4 vs FP16 on other cards.
DegustatoR
Legend
So far nobody is using DirectML for any upscaling. No idea why but it seems weird enough to make an assumption that it's not fit for the task. Also - what would you use for that on non-Windows platforms, in games using Vulkan?
DirectSR just means that the upscaler is shipped with the driver instead of the end s/w. It doesn't change anything in API limitations and is in fact more limiting than what we have now.
DirectSR just means that the upscaler is shipped with the driver instead of the end s/w. It doesn't change anything in API limitations and is in fact more limiting than what we have now.
DavidGraham
Veteran
In their talks about XeSS, Intel mentioned DirectML was too slow for them to use it to implement XeSS, even for XeSS version that runs on hardware other than Intel.
XeSS isn't vendor exclusive, only the better model is.
In fact this is probably why XeSS is built like it is: use the generic hardware-agnostic-ish (there are requirements but they are fairly broad) to get developers to implement the API, the Intel exclusive model gets to piggy back off this implementation.
If FSR4 was RDNA4 exclusive I would imagine they would have to craft it so games that already have implemented FSR2/3 don't have to do much reworking (or its a driver-side 'swap' essentially, apologies for the lack of technical terminology here lol im out of my depth). Essentially anyone who isn't Nvidia needs to use the hardware-agnostic version as a 'trojan horse' (I don't mean this in a derogatory way like people usually say trojan horse but I can't think of another analogy here lol) for their hardware-exclusive version.
It’s no different to FSR where there may be vendor agnostic shader based (FSR 3) and vendor locked ML based (FSR 4) models. From an application perspective it’s likely the same API call to use either.
Yes this is my point, thank you for wording it better lol
Intel cards are really bad with directML. AMD and Nvidia are not so far off the mark. Intel will need to do something if they want to get DML performance up.