Colfax and MIT speed up lattice QCD staple networks 700–1200x on H200 with CuTe DSL

snowclipsed · x · 2026-11-21

Colfax Research, collaborating with MIT physicists, published a blog detailing optimizations to the staple network—a core component of normalizing flows for lattice QCD Monte Carlo sampling. On an NVIDIA H200 NVL, their CuTe DSL kernels accelerate forward-pass hidden layers by 700–1200x over a reference PyTorch implementation and 9–11x over a previous fused Triton version, and 2.0–3.4x over a Triton port of their optimizations. The workload—many small 3x3 complex double-precision matrix multiplies in a stencil pattern on a 4D lattice—differs sharply from typical LLM operators; the post covers the physics background, per-layer math, kernel design, and further optimizations for more channels.

Original post →

More from Infra

Infra channel →