CUDA version causes 3.3x speed difference in quantized video models; B200 loses to properly configured 4090
Odd_Lavishness2236 · reddit · 2026-08-14
A Reddit user shares benchmarks migrating LTX-2.5 video model. On the same RTX 4090, upgrading PyTorch from 2.8.0+cu129 to 2.13.0+cu130 yields 3.3x speedup (89.4s vs 27.1s per clip). The reason: below cu130, ConvRot weights are upcast to bf16, negating quantization. Also, ComfyUI's --fast fp8matrixmult flag significantly helps datacenter cards like B200 (58.0s to 28.1s). The author warns to check torch version before benchmarking, as it can invalidate GPU comparisons.
More from Infra
- $1T of AI Infra Buildout to Solve $1M Math Problems? — suchenzang · 2026-08-14
- Intel and MiniMax Release MXFP4 Quantized Model, Topping FP4 Benchmarks — HaihaoShen · 2026-08-14
- GPU Prices Surge Again: RTX 6000 Pro Jumps to $15k — HankYeomans · 2026-08-14
- Building a Multi-GPU Workstation for Local 122B LLM Inference on a Budget — whatyathinkk · 2026-08-14
- No Real GPU Shortage, But Rental Terms Are Brutal: 3-5 Year Commits — AccBalanced · 2026-08-14
- Cloud Workstations Reshape Embodied AI R&D Infrastructure — 量子位 · 2026-08-14