Discussion RDNA 5 / UDNA (CDNA Next) speculation

[Kepler_L2]

They can't run on base PS5?

Work Graphs is only supported on RDNA3 and later, and is really only optimized for performance in RDNA5 and later.

 

[Win2012R2]

Work Graphs is only supported on RDNA3 and later, and is really only optimized for performance in RDNA5 and later.

Ah, that's a pity, but at least this looks like nice exciting stuff for next gen - far more promising than mesh stuff and RTRT obviously limited by hardware

 

[Magras00]

Seems like work graphs got a lot more potential benefits

100%. It's a bigger deal than DX12 and Vulkan but it's a new paradigm that requires re-educating engine SWE and complete rewrite of engine components. Will take a while but everyone should switch when crossgen is done. It's just much easier to work with than Execute Indirect (EI). GDC 2024 video explains the benefits:

Work Graphs is only supported on RDNA3 and later, and is really only optimized for performance in RDNA5 and later.

Still no PS5 Pro support? Stupid question of course it doesn't. RDNA 2 binaries. Custom ML and RT only things beyond RDNA .

@Kepler_L2 are these HW optimizations in GFX13 for Work graphs?

1. SWC (reordering)
2. WGS (Shader Engine level)
3. Local launchers (WGP level)?
4. Anything else I missed?

AMD already touted work graphs vs IE = 39% less ms for PCG Procedural content generation with 7900XTX and you're tell me it isn't even properly optimized for performance at HW level. Impressive!
Large gap in IPC gain (EI vs Work graphs) for RDNA 4 vs RDNA 5 expected With GPU Workgraphs we should expect a significantly larger IPC gain from RDNA 4 -> 5 compared to the gain with EI. It sounds like launch IPC testing def won't be indicative of RDNA 5's full potential.

For Work graphs vs EI does a doubling of RDNA 4 -> RDNA 5 raster IPC gain sound too high?

Edit: Retracted unrealistic claims.

 

[soresu]

AMD already touted work graphs vs IE = 39% less ms for PCG with 7900XTX and you're tell me it isn't even optimized for performance at HW level. Impressive!

Could just as easily mean that the current sw solution(s) that work graphs are designed to replace are pretty suboptimal, such that even a relatively unoptimised hw µArch can extract a decent perf uptick.

Vulkan is in a constant state of revision as the various hw and sw stakeholders make their bids to extend it, make it easier to use or optimise.

It wouldn't surprise me at all to learn that certain features of the API are essentially just lower level equivalents to what was already in OpenGL or DirectX at the time of Vulkan's inception and hadn't yet been replaced with a more optimal solution.