Softwired Technologies

We are happy to sponsor the new, open-source Ladybird web browser! We believe in a truly open web. Supporting projects like Ladybird helps ensure the web remains free, accessible, and independent for everyone. https://lnkd.in/d4DCiAuE

Tensor cores in #webgpu WebGPU has a new feature called subgroup matrix, currently behind a Chromium experimental flag. It exposes Metal's simdgroup functionality (simdgroup_float8x8, simdgroup_load, simdgroup_multiply_accumulate, etc.), and maps to similar constructs in backends such as Vulkan or DirectX. Subgroup matrix allows WebGPU shaders to express warp-level matrix math. I created a new Dawn build from master that has this feature, and integrated it into pydawn, then wrote a matmul kernel that: - shows how to define lhs, rhs and result matrices (subgroup_matrix_{left/right/result}<{datatype}, 8, 8>) - features subgroupMatrixLoad to load matrices from memory - shows how to calculate offsets in a tiled matmul - shows how to store a subgroup matrix into memory And the good part is that it is all in Python, thanks to the autogenerated pydawn bindings: https://lnkd.in/dNXW9Ebx

brainchop.org now runs on #tinygrad #WebGPU - a project we’ve been collaborating on.

Our first blog post dives deep into the internals of tinygrad’s WebGPU backend. WGSL doesn’t support sub-32-bit integer types, so we had to get creative - the solution is bitpacking. If you’re ready to revisit bitshifts, masks, and some low-level GPU fun, this one’s for you. https://lnkd.in/dzkR97tG #webgpu #tinygrad #gpgpu #bitpacking #atomics