Modders unlock Nvidia's half-speed AI throttle on GeForce RTX 20, 30 and 40-series cards

Open-source Linux kernel modules (gpu-unlock, gpu-unlock-ada and turing-unlock) remove the cap Nvidia places on FP16 and BF16 tensor math with FP32 accumulation on consumer GeForce cards, which runs at half the rate of its pro cards. The tools exploit a stack overflow in Nvidia's signed SEC2 booter firmware to rewrite two speed-select registers before the stock driver loads, with no VBIOS flash or patched driver. Peak tensor throughput on the RTX 4090 doubles from 165.6 to 331.6 TFLOPS, but real matrix-multiply gains in one independent test were 12–34%. Ampere users report LLM prompt processing up 14–36% and faster image generation, while token generation, INT8 and gaming are essentially unchanged. The unlock is experimental and risks boot failures.