Linux kernel module unlocks higher prefill performance on older Nvidia consumer GPUs

A new Linux kernel module, gpu-unlock, significantly improves prefill performance on older Nvidia consumer graphics cards by overriding restrictive driver settings. The patch targets Nvidia 3000-series GPUs, including the RTX 3090, 3080, and 3070, by modifying the pre-driver timing of the UEFI interface. By setting the speed selector to 0x88888888, the module enables the full BF16 instruction rate, effectively bridging the performance gap between consumer hardware and professional-grade equivalents. Benchmarks using the Qwen 3.8 27B model demonstrate substantial gains in prefill tokens per second across various context lengths, with the 2K context length showing an increase from 1157 to 1840 tokens per second. The module is compatible with Proxmox and other Linux distributions, requiring root access and a kernel command-line flag to activate. While the performance gains are not strictly double, they provide a significant boost for users running local AI workloads on consumer hardware. The developer notes that the patch does not require manual updates when Nvidia releases new drivers, though it is not a replacement for a certified professional driver.

The gpu-unlock kernel module enables the full BF16 instruction rate on Nvidia 3000-series consumer GPUs. Performance testing with the Qwen 3.8 27B model showed a 59% increase in prefill speed at a 2K context length.

The patch functions by overriding the speed selector in the UEFI interface to bypass artificial performance limitations. Users can apply the patch by adding the 'gpu_unlock.targets=all' flag to their kernel command line.

The module is compatible with various Linux distributions and virtualization platforms like Proxmox. The developer suggests that the RTX 2080 Ti, when modified with 22GB of VRAM, may offer the best price-to-performance ratio for local AI tasks.