FP8 + Ozaki Scheme II delivers 500 TFLOPS FP64 on B300, says new paper

A bold new paper from renowned HPC expert Satoshi Matsuoka (Tokyo Tech / Riken) challenges the long-held dogma that native FP64 hardware is essential for scientific computing. Titled "FP8 is All You Need (Part 1): Debunking Hardware FP64 as the HPC Holy Grail", the paper shows that on AI-optimized GPUs like NVIDIA's Blackwell Ultra (B300), using abundant FP8 tensor core throughput combined with the Chinese Remainder Theorem-based Ozaki Scheme II can recover full FP64 accuracy across a spectrum of HPC kernels. The B300's native FP64 performance has collapsed to just ~1.3 TFLOPS—a 31x regression from the B200—making even memory-bound kernels become compute-bound. Matsuoka introduces the Tensor-Memory Equilibrium (TME) model, augmenting the Roofline model with a compute multiplier (alpha), bandwidth multiplier (beta), and reconstruction latency (gamma), and identifies register-level fusion as the key mechanism driving beta near 1, effectively making emulation free behind the memory wall.

Results show that Ozaki II vaults emulated FP64 from the ~1 TFLOPS native floor to ~500 TFLOPS on the B300 and ~400 TFLOPS on AMD's upcoming Rubin R200, exceeding even B200's native FP64 ceiling by over an order of magnitude in the compute-bound regime while matching the memory roof in bandwidth-bound regimes. Against an H100 baseline, Ozaki II matches or exceeds performance on every workload studied, versus the up-to-50x regression that B300 native FP64 imposes. A companion Part 2 paper covers FFT analysis using Kulisch fixed-point reconstruction on the INT32 pipe and FP32+Kahan reductions. The evidence strongly supports the title's claim: FP8, with Ozaki II and Kulisch escape routes, is sufficient for production HPC—native FP64 silicon is no longer the holy grail it was once thought to be.