GPU unlock 1.0.0 is a Linux kernel module that unlocks the FP/BF16 instruction issue rate on Nvidia 3000 series cards to match their professional equivalents. The module combines a CMP unlocker V67/PLM technique with pre-driver UEFI timing. It runs the unlock, writes speed selectors and performs FLR before the standard NVIDIA driver binds, eliminating the need for manual patching with new driver releases.
The effect doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations on GA102, GA103, GA104 and GA106 architectures. Testing on RTX 3090 Ti shows BF16 GEMM throughput increasing from 87.50 to 157.41 TFLOPS, while 3090 models see gains of 25 to 67 percent. Supported cards include RTX 3090 Ti, 3090, 3080 Ti 20GB OEM, 3080 10GB and 3070, with gains reported up to 80 percent on RTX 3060 12GB.
The module accepts all GeForce IDs listed in the source code. Matrix-heavy workloads see improvements, with LLM prefill speeds increasing 14 to 36 percent. LLM decode and INT8 work remain bandwidth-limited and largely unchanged.
Users report bit-identical outputs before and after the unlock. The module supports multi-GPU setups, display-driven cards and VM passthrough via vfio-pci. Installation requires appending gpu_unlock.targets=all to the kernel command line and building the module from source.
Source: Hacker News · Summarized by HeadlinesBriefing