FP8 Tensor Cores Enable FP64-Equivalent 3-D FFT on NVIDIA Blackwell Ultra via Ozaki-Bailey Decomposition

Summary

This technical paper presents the Ozaki-Bailey FFT, a method for achieving full FP64-equivalent 3-D FFT performance on NVIDIA's Blackwell Ultra (B300) GPU, which has severely reduced native FP64 throughput (~1.3 TFLOPS, ~30x below B200). The approach uses the Ozaki Scheme II framework to route dense matrix multiply through FP8 tensor cores with mantissa-sliced Chinese-remainder reconstruction. The FFT implementation uses Bailey's six-step decomposition with both 1-D FFT GEMMs on FP8 tensor cores. Garner reconstruction splits into Phase A (inner products on FP8/INT8 tensor cores) and Phase B (per-output reduction), where the authors identify Kulisch fixed-point complete arithmetic as a reformulation that maintains full FP64 accuracy on the INT32 SIMT pipe. The paper derives bandwidth-parity floors and projects ~18 ms for 1024^3 at full FP64, approaching the 12.9 ms memory roof. If the projection holds, B300 becomes viable for full-FP64 FFT through software alone, motivating a libKulisch library and benchmark campaign.

Source

bskyFP8 Tensor Cores Enable FP64-Equivalent 3-D FFT on NVIDIA Blackwell Ultra via Ozaki-Bailey Decompositionarxiv.org