NVIDIA GeForce RTX GPUs see massive AI performance boosts via Linux mod

Open-source developers have released tools that bypass software-imposed limitations on NVIDIA GeForce RTX 20, 30, and 40 series graphics cards, significantly increasing AI computing performance. The projects, including ‘gpu-unlock’ and ‘GPU Unlock Ada’, utilize Linux kernel modules to modify low-level hardware registers before the official NVIDIA driver loads. This method avoids the need to flash the VBIOS or patch drivers.

The modification targets the instruction issue rate for Tensor Core operations, specifically for FP16 and BF16 formats with FP32 accumulation. In consumer cards, these rates are often halved compared to professional workstation models. Testing shows dramatic theoretical gains: the RTX 4090 can see FP16 throughput jump from 165.6 TFLOPS to 331.6 TFLOPS, while the RTX 3090 Ti shows an approximately 80% increase in BF16 performance.

While these boosts are substantial for synthetic benchmarks and specific AI workloads like image generation and LLM prefill stages, they do not translate to doubled gaming frame rates. Real-world AI performance gains are more modest due to memory bandwidth bottlenecks during the decoding phase. The modifications are temporary and reset upon power loss or system reboot.