Open-Source Tool Unleashes Nearly 80% Boost in BF16 Compute Power for RTX 30 Series GPUs
Zagitta, a developer, has created an open-source tool called gpu-unlock that removes the instruction rate limit on FP16/BF16 Tensor Cores for RTX 30-series GPUs, without requiring VBIOS or driver modifications. In tests, the BF16 computing power of an RTX 3090 Ti increased from 87.5 to 157.4 TFLOPS, a rise of nearly 80%. The actual performance gain varies depending on the workload: local LLM generation speed remained almost unchanged, while the impact on game FPS has yet to be independently…
Open-Source Tool Unlocks Nearly 80% Boost in BF16 Compute Power for RTX 30 Series GPUs
The open-source tool gpu-unlock, developed by developer Zagitta, can remove the instruction rate limits on the FP16/BF16 Tensor Cores of RTX 30 series graphics cards, without requiring any VBIOS or driver modifications. In tests, the BF16 compute power of an RTX 3090 Ti increased from 87.5 to 157.4 TFLOPS, a gain of nearly 80%.
However, the actual performance improvement depends on the workload: local LLM generation speeds saw almost no change, and the FPS boost in games has not yet been independently verified. The tool currently has primary support for Linux and remains experimental.
Sources: Hong Kong Mobile Game Network GameApps.hk | GitHub