Open source developers have released an experimental tool called GPU Unlock that doubles artificial intelligence computing speeds on consumer NVIDIA graphics cards. The software bypasses artificial firmware limits on Ampere, Ada Lovelace, and Turing architectures without altering official drivers or flashing the card VBIOS. Benchmarks show tensor processing jumps by up to 80% on certain hardware, though gaming performance remains unchanged.
Zagitta developed GPU Unlock by targeting the instruction issuing limits that NVIDIA places on consumer silicon. Consumer cards share identical dies with enterprise products, but firmware restricts the execution speed of FP16 and BF16 tensor core math using FP32 accumulation. The tool injects a custom Linux kernel module during the boot sequence before the graphics driver initializes. This module updates the hardware speed registers, triggers a PCIe Function Level Reset, and allows the standard driver to load normally. A separate experimental setup runs through UEFI on Windows machines, though compatibility is currently restricted to select cards.
Testing demonstrates massive computing speedups across multiple GPU families. On an RTX 4090 running at 450W, FP16 tensor throughput surged from 165.6 TFLOPS to 331.6 TFLOPS, matching the metrics of enterprise grade hardware like the RTX 6000. An RTX 3090 Ti at 450W gained an 80% speedup in BF16 matrix multiplication, which directly cut generation times in ComfyUI image workflows by roughly 58%. An RTX 3080 Ti saw BF16 throughput jump from 71.59 TFLOPS to 113.98 TFLOPS at 315W, while RTX 20 series Turing cards recorded similar 2x compute improvements.
Real software applications reveal clear performance divides. Local image synthesis tools profit immediately from the unlocked math rates, yet large language models show mixed outcomes. An RTX 3090 testing vLLM prompt processing achieved a 33% throughput increase during the prompt phase, while output token speeds showed zero change. Memory bandwidth bottlenecks the text decoding phase rather than raw tensor core speed. The software cannot increase memory bandwidth, activate disabled physical CUDA cores, or boost DLSS frame rates in video games.
