On Oct. 31, VideoCardz reporter WhyCry reported on physicsdolphin’s NV UEFI Unlocker, a new UEFI application that brings earlier Tensor Core unlock techniques into one tool, including for Windows systems. It is designed to change GPU behavior before Nvidia’s graphics driver loads. It targets certain Tensor Core operations on GeForce RTX 20-, 30- and 40-series GPUs. That covers tested models across three generations, not every card in those families.
WhyCry explains that the utility boots from a prepared USB drive, applies register changes and then hands control to the operating system. It does not permanently modify the GPU’s firmware, but the unlock must be reapplied after a cold boot.
What the benchmarks show
Wccftech reporter Sebastian Castellanos reports that an RTX 2070 went from 36 to about 60–63 TFLOPS in FP32-accumulation matrix multiplication, a 67–75% increase. A separate RTX 4060 Laptop GPU FP16 matrix-multiplication test rose from 29.89 to 49.33 TFLOPS, about 65%, according to the project’s account. These are results for particular devices and tests, not expected gains across AI applications.
VideoCardz notes that gains depend on the workload. As an example, it cites a user-reported 28.8% improvement in ComfyUI SDXL on an RTX 3080 LHR, smaller than the synthetic benchmark gains.
The standard setup requires disabling Secure Boot, VideoCardz reports. RTX 20-series GPUs also need an additional GSP firmware configuration step. The outlet identifies the RTX 2070, RTX 2060 SUPER, RTX 3060 Ti LHR and RTX 4060 Laptop as verified examples, while noting that not every GPU and system configuration has been independently tested.
Comments