10 ms·
The GPU people are also reaching for simd and fixed matmul hw to increase perf. Tensor Cores (int, fp16, tf32 and even fp64 on A100) and the new DPX instruction
by touisteur 4y ago
The GPU people are also reaching for simd and fixed matmul hw to increase perf. Tensor Cores (int, fp16, tf32 and even fp64 on A100) and the new DPX instructions. RT cores are a different kind of horse but still specialized for BVH traversal and ray-triangle intersection.
- theLiminator 4y agoWe're reaching a point where CPUs are increasingly getting more specialized, and GPUs are becoming increasingly generalized. Going at improvements from both sides of the sandwich.