user Zagitta has released gpu-unlock, an open-source tool designed to bypass performance restrictions on RTX 30-series GPUs. By targeting hardware speed-selection registers via a kernel module, the tool unlocks restricted Tensor Cores.
The modification operates without flashing VBIOS or using patched drivers by executing before the Nvidia driver loads. Benchmarks show an increasing BF16 throughput from 87.5 to 157.4 TFLOPS, while ComfyUI tests saw up to 59% higher throughput.
While local LLM inference shows improved prompt processing, token generation remains unchanged and gaming gains are unverified. Originally for Linux and based on CMP mining research, an experimental Windows port of the Ampere-focused tool is now available.