GoKawiilOpen-source kernel mods unlock full-speed FP32-accumulate AI math on GeForce RTX 20/30/40 GPUs
Developers have released Linux kernel modules—gpu-unlock, gpu-unlock-ada, and turing-unlock—that rewrite speed-select registers on Nvidia RTX 20-, 30-, and 40-series GPUs before the driver loads, removing a hardware-level cap on FP32-accumulate tensor math that Nvidia normally reserves for workstation cards. Benchmarks show gains ranging from roughly 12% to nearly 2x depending on card and workload, with an RTX 3090 Ti jumping from 87.5 to 157.4 BF16 TFLOPS and an RTX 4090 doubling its raw FP16 instruction rate, though real-world matrix workload gains are smaller.