RTX 3090 ships half-speed fp32 tensor math; gpu-unlock module boosts BF16 throughput 1.9x
sn2006gy · reddit · 2026-10-22
A Reddit thread highlights developer Alexey Fateev's finding that NVIDIA ships every RTX 3090 with fp32-accumulate tensor math running at half speed (the driver even reports FMLA32 = 1/2). The gpu-unlock kernel module by Zagitta sets it to full speed before the NVIDIA driver loads.
Measured results on the same cards, driver, and 300W power limit:
- BF16 tensor pipe: 82 → 155 TFLOPS; cuBLAS BF16 GEMM at 8192: 64 → 99 TFLOPS; TF32: 41 → 79 TFLOPS.
- Real inference gains are smaller: prefill of Qwen3.8 Flash-Next on two cards via Strata improved 10-12% across 4K/10K/90K contexts, since GEMM phases are only 40% of Strata prefill.
- Outputs are byte-for-byte identical to stock runs, confirming a pure throughput unlock.
The thread also wonders whether such caps trace back to anti-mining hacks and if other GPUs could benefit.
More from Infra
- Crusoe unveils Tinker API-compatible LLM post-training platform at PyTorch Conference — PyTorch · 2026-10-22
- Socher's Recursive launches Flint: local coding model at 125+ tok/s on MacBook — ying11231 · 2026-10-22
- SGLang-TorchTPU launches with Google and Meta, hits 770 tok/s decode in production — ying11231 · 2026-10-22
- Open PR optimizes DeepSeek-V4.1 inference on RTX PRO 6000 Blackwell — AlpinDale · 2026-10-22
- vLLM adds day-0 support for Flint: 500 TPS on a single GB200 — _rockt · 2026-10-22
- Grass founder on the Delphi Podcast: training data is now a profitable business serving top labs — grass · 2026-10-22