📅 October 31, 2026 • VideoCardz
An open-source tool called nv-uefi-unlocker, developed by a creator known as physicsdolphin, removes the Tensor Core cap on NVIDIA GeForce RTX 20, 30 and 40 series GPUs for AI workloads. The tool applies firmware-level register changes from a USB boot before the NVIDIA driver loads. No VBIOS flash is required, the changes work on Windows, and they reset after a cold boot, which makes the modification temporary by design.
Background
NVIDIA has long limited certain compute paths on consumer GeForce cards to separate them from data-center hardware. These caps affect the precision formats widely used in AI, such as FP16 matrix operations. Community modders have explored ways around the limits for years, and nv-uefi-unlocker automates the approach so it no longer requires manual register editing.
Verified gains
Community testing has produced concrete numbers. On an RTX 4060 Laptop, FP16 GEMM throughput rose from 29.89 to 49.33 TFLOPS, a gain of about 65 percent. RTX 2070 owners reported 67 to 75 percent higher FP32 GEMM performance, and earlier community tools had shown gains of 58 to 85 percent on RTX 30 series cards. Gaming FP32 performance, by contrast, is unchanged. The figures come from users running standardized GEMM benchmarks before and after applying the register changes.
Real-world caveats
Gains in actual AI workloads are smaller than the raw benchmark numbers suggest. One RTX 3080 LHR user reported a 28.8 percent speedup in ComfyUI SDXL image generation. The tool does not support RTX 50 Blackwell cards, and support for GTX 16 series remains experimental. Because the register changes are volatile, they must be reapplied after every cold boot.
How it differs from flashing
Unlike a VBIOS flash, which permanently rewrites firmware and carries a real risk of damaging a card, the USB-boot approach changes registers only for the current session and leaves the stored firmware untouched. That lowers the risk of permanent harm, though users should still be aware that modified operation may fall outside vendor support policies.
Impact
For budget-conscious AI builders, the tool offers a way to recover Tensor Core performance that consumer cards leave disabled, most visibly in FP16 and FP32 compute used in image generation and local inference. The gap between microbenchmark gains of up to 85 percent and the roughly 29 percent observed in a real ComfyUI workflow shows that end-to-end speedups depend on the workload, not just the hardware. The volatility of the change also means clusters and shared machines behave predictably after a restart, returning to stock behavior without user intervention.
Sources: videocardz.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)