By Peter Paul
Publication Date: 2026-10-25 00:23:00
A new open-source GPU unlock mod has pushed NVIDIA’s GeForce RTX 3090 Ti to nearly 80% higher BF16 compute performance, with some real-world AI workloads gaining almost 60% in throughput. The modification also works on several other RTX 30-series graphics cards, including the RTX 3060 and RTX 3070. Separate experiments have now brought similar performance unlocks to the RTX 4090 and RTX 4070 Ti SUPER.
GPU Unlock Removes RTX 30-Series Tensor Core Restrictions
Developed by GitHub user Zagitta, GPU Unlock is an open-source Linux kernel module that enables higher Tensor Core instruction-issue rates on supported NVIDIA Ampere graphics cards.
The project targets a compute restriction affecting FP16 and BF16 operations with FP32 accumulation. These operations normally run at a reduced instruction-issue rate on affected GeForce GPUs compared with their professional counterparts.
GPU Unlock changes the relevant speed selectors before NVIDIA’s driver initializes the graphics card. It combines techniques from earlier CMP GPU unlocking research with an early-boot implementation, allowing the regular NVIDIA driver to load afterward.
The modification doesn’t require a patched NVIDIA driver or a permanent VBIOS flash. It can also continue…
