H100 Memory Study Gets Errata After DeepSeek's Open-Source V3 Trace Reveals SwiGLU Quantization Fusion Details

ezyang · x · 2026-09-02

PyTorch core dev ezyang confirmed a fix and errata in the H100 memory study after Vlad Savinov flagged an issue: SwiGLU doesn't consume e4m3 directly since DeepGEMM outputs bf16 — quantization is likely fused into the SwiGLU kernel.

Vlad backed this up with DeepSeek's open-sourced V3 perfetto trace (the profile-data repo on GitHub, 1.2k stars): SwiGLU time matches topKseqlenintermediate7 / 2.8e12 ≈ 0.16ms. He infers the kernel reads two bf16 tensors and outputs three fp8 tensors — two quantized inputs plus a quantized silu(a)b feeding the next DeepGEMM call (2.8e12 is 85% of Hopper's 3.35 TB/s spec bandwidth).

Original post →

More from Infra

Infra channel →