On October 25, 2026, Tensorust reported that an experimental gpu-unlock method can raise FP16 and BF16 tensor throughput on GeForce RTX 30-series cards, with the upside aimed at AI math rather than general GPU acceleration.
According to Tensorust’s review, the Linux-first approach runs before Nvidia’s driver loads and changes hardware speed-selection registers, without flashing VBIOS or patching the driver. Tensorust’s roundup of 18 public reports covering 24 cards found a median BF16 gain of 66%, with results ranging from 22% to 99%. One RTX 3090 example went from 76.9 to 127.4 TFLOPS.
The gains are not uniform across AI work. Tensorust and other coverage say tensor-heavy stages such as prompt prefill, fine-tuning, and some image-generation workloads improve more, while token decode and other paths may show little change. Reporting also describes the method as experimental, with the clearest support on Linux, limited unofficial Windows use, and no solid evidence of gaming gains.