[HN Gopher] Cooperative Vectors Introduction
___________________________________________________________________
Cooperative Vectors Introduction
Author : JasperBekkers
Score : 41 points
Date : 2026-04-10 13:16 UTC (2 days ago)
(HTM) web link (www.evolvebenchmark.com)
(TXT) w3m dump (www.evolvebenchmark.com)
| cubefox wrote:
| They talk about three applications for the coming cooperative
| vectors / linear algebra API: Neural Materials, Neural Radiance
| Caching, Neural Texture Compression.
|
| But what about already much more popular ML-based applications:
| upscaling, frame generation, denoising? These are currently
| vendor specific. E.g. DLSS is only supported on Nvidia GPUs.
| Could these run on arbitrary DirectX 13 GPUs in the future by
| using the new API?
|
| Granted, GPU manufacturers would likely not be keen on getting
| their tech on other platforms, but engine developers like Epic
| would clearly want to use it if they come up with their own ML
| upscalers, denoisers etc in the future.
| atq2119 wrote:
| It's sort of explained in the article, actually.
|
| Upscaling and related applications are already supported by
| cooperate _matrix_ ops. Those shaders /kernels run in
| workgroups that operate on a tile of pixels at a time, so you
| naturally get matrix-matrix operations. And yes, at least in
| Vulkan, there has been cross-vendor support for those for a
| while now.
|
| Neural materials and similar applications often operate on a
| single datum at a time, like in a pixel shader or a raytracing
| shader.
|
| Yes, the hardware still groups the threads of a pixel shader
| together, but there is inherently more divergence. And so the
| natural program representation tends to have matrix-vector
| operations instead of matrix-matrix operations.
|
| That's where cooperative vectors come in.
___________________________________________________________________
(page generated 2026-04-12 23:01 UTC)