Pure Rust GEMM on Nvidia Blackwell, part 2 🦀
Part 1 broke down the kernel. Part 2: measurements.
48-SM Blackwell, vs live cuBLASLt FP16:
4096³ ▸ 445 TFLOPS (98.7%)
8192³ ▸ 518 TFLOPS (97.8%)
16384³ ▸ 507 TFLOPS (119% 🤯)
We're not SoL, sometimes we're past it. 🧵👇