Claude-written CUDA C++ port runs TRELLIS.2 5.3x faster on H200, bit-identical
huggingface · x · 2026-11-17
A developer had Claude rewrite the TRELLIS.2 image-to-3D pipeline fully in CUDA C++, achieving 5.3x speedup on H200 over the original PyTorch implementation with bitwise parity by default (the only exception being vertex normals, which are nondeterministic in the official code too).
- Surprising finding: the bottleneck wasn't the neural net but the GLB export function — moving export to the GPU delivered the biggest gains
- Zero Python at runtime; total cost $7 in Claude usage plus $60 of H200 compute via Hugging Face Jobs
- Open-sourced as trellis2-cuda (Apache-2.0) on GitHub with a full worklog (docs/blog.md); a tolerance-bounded fast mode is available via T2FAST=1
More from Infra
- Fireworks AI tells investors chips are a commodity, its moat is software that runs models fast — hargup13 · 2026-11-17
- China to make 3M AI ASICs in 2027 vs West's 20M, each ~4x faster per chip — teortaxesTex · 2026-11-17
- Harmonic's AI theorem prover Aristotle takes a no-hallucination hardware synthesis talk to Cambridge — satnam6502 · 2026-11-17
- GPU compute futures pushed past Jan 8 as commodity market takes shape — YiMaTweets · 2026-11-17
- Worst-case: a Carrington-class solar storm could wipe out million-satellite orbital data centers — johnseach · 2026-11-17
- Memory spot prices keep falling: DDR5 RDIMM down ~10% in 2 weeks, NAND -33% — SumitGup · 2026-11-17