Linux GPU-unlock module aims to speed up matrix math on some RTX 30-series cards

Zagitta's GitHub README describes a Linux module that changes speed selectors before Nvidia's unmodified driver binds certain RTX 30-series cards. It lists user-reported BF16 matrix-multiplication gains of 25% to 67% for RTX 3090 cards, but says the reports use different benchmarks and power limits affect results. The project says matrix-heavy work benefits while LLM decode and INT8 work change little; enabling the module requires a boot option.