ThunderKittens' NVFP4 GEMM on Vera Rubin went from 42% of roofline to over 22 PFLOPS
Together AI Blog·low signal
Together ported ThunderKittens to Vera Rubin NVL72. Running its Blackwell GEMM unchanged reached about 42% of roofline, and a rebuilt NVFP4 kernel reached over 22 PFLOPS, which Together says is competitive with cuBLAS and CuTe DSL. The gains came from the new ISA: a doubled K-step width, 576-column tensor memory, 328 KiB shared memory, B-side collectors and early A-operand release in PTX 9.4. It shows that Blackwell kernels leave half of Rubin unused without a retune.