← All resources

Adventures in Neural Rendering part 2: Cooperative vectors

Auto-discovered

Blog Advanced

Kostas Anagnostou

Visit resource ↗

Description

Technical blog post by Kostas Anagnostou on Interplay of Light, part 2 of Adventures in Neural Rendering, on Cooperative Vectors for MLP inference in shaders. This URL is one deep-dive article.

It explains Tensor-core MMA, mapping MLP layers to matrix ops, and enabling Cooperative Vectors with a preview Agility SDK, SM6.9 DXC, experimental features, and float16 weights in ByteAddressBuffers. It covers MulAdd, alignment, leaky ReLU, and speedups versus float32 compute inference for radiance and RTAO-style nets on an RTX 3080 mobile.

Readers need Direct3D 12, HLSL, and basic neural matrix math. Assumes the prior series part; targets advanced graphics engineers.

Free WordPress article with code, feature checks, and traces. Notes this Cooperative Vectors preview is later superseded by a Linear Algebra Matrix path.

Useful for judging when hardware-accelerated HLSL inference beats a hand-written compute MLP.