The open-source gpu-unlock project uses a Linux kernel module loaded before NVIDIA’s unmodified driver to relax FP16 and BF16 tensor-operation limits with FP32 accumulation on selected GA102, GA103, GA104 and GA106 GPUs. Its RTX 3090 Ti benchmark increased BF16 throughput from about 87.5 to 157.4 TFLOPS. A separate RTX 3090 user test reported roughly 14% faster long-prompt processing, while memory-bound token generation barely changed. Results depend on the card, power settings and workload; they do not establish universal twofold speedups or long-term stability.
Back to Tech
Tech
gpu-unlock raises tensor throughput on some Ampere GPUs
An early-boot Linux module improves selected operations, not every stage of LLM inference.
Sources
- Zagitta/gpu-unlock: Unlock FP/BF16 instruction issue rate on Nvidia 3000's cards to their professional equivalents , GitHub
- Compatibility report: stock EVGA RTX 3090 (GA102, 10de:2204) — works on Ubuntu 24.04 / kernel 7.0.0-38 / driver 610.57.04-open; ~2x FP16/BF16 per-MHz under a 350 W cap; warm reboot re-unlocks (selectors read 0)