Linux Kernel Module Unlocks Half-Speed Tensor Cores on RTX 3090, Boosting BF16 to 1.9x
UltramanQuar · reddit · 2026-10-23
A Reddit user relays and probes a claim from X: NVIDIA ships every RTX 3090 with fp32-accumulate tensor math at half speed, and the driver even reports it as FMLA32 = 1/2. Zagitta's gpu-unlock project sets it to full before the NVIDIA driver loads; after a cold boot all four cards report FMLA32 = full.
Measured results (author's own runs):
- BF16 tensor pipe: 82 → 155 TFLOPS
- cuBLAS BF16 GEMM 8192: 64 → 99 TFLOPS
- TF32: 41 → 79 TFLOPS
- Real inference (Strata, Qwen3.8 Flash-Next, 2 cards): +12% prefill at 4K/10K, 90K from 3205 to 3536 tok/s (+10%), below the predicted +14–19%
The author notes GEMM phases got 23% shorter but are only 40% of Strata prefill, while int8 experts, copies and CUDA core kernels saw no speedup. Outputs are unchanged: with the same expert cache size, the log-prob dump over 1996 teacher-forced positions is byte-for-byte identical to the stock run. The author asks whether 40-series cards benefit too.
Related event: Community patch unlocks RTX 3090 tensor performance, boosting BF16 by 1.9x(3 posts)→
More from Infra
- Samsung Electro-Mechanics joins Nvidia's substrate supply chain, shocking Taiwanese rivals — zephyr_z9 · 2026-10-23
- Vercel founder Suhail publicly offering to buy RL kernel optimisation tasks for models — AIFlow_ML · 2026-10-23
- GitHub moved 99.99999% of repos to Azure; per-reference quorums unlock agent-scale Git concurrency — marlene_zw · 2026-10-23
- Jensen Huang visits Astana, joins Kazakhstan's AI minister to discuss nation's AI future — nvidia · 2026-10-23
- FP16 vs FP32 matmul on RTX 4060: 3.5x speedup at large sizes, slower at 128x128 — cneuralnetwork · 2026-10-23
- BaseQuant V1 quants for agentic Qwen3.8-27B hit 17-34% lower KL than Unsloth — nightowlsleeping · 2026-10-23