Open-source GPU Unlock lifts tensor core limits on NVIDIA RTX 30 series, boosting performance by over 60%

A GitHub developer known as Zagitta has released GPU Unlock, an open-source project that removes a performance limiter on NVIDIA's Ampere-based GeForce RTX 30 series graphics cards. The tool uses a Linux kernel module to lift an instruction issue rate cap on tensor core operations, potentially delivering more than 60% higher performance in some workloads.

The project targets a specific limitation where consumer Ampere GPUs run FP16/BF16 tensor core operations with FP32 accumulate at lower rates than professional cards. It modifies a hardware speed-select register and performs a function-level reset before the NVIDIA driver loads, avoiding permanent VBIOS flashing or driver patches. Supported GPUs include GA102, GA103, GA104 and GA106.

Benchmarks shared by the developer and community show notable gains: an RTX 3090 Ti at 450W achieved significant improvements in dense GEMM tests; an RTX 3080 Ti saw BF16 throughput rise from 71.59 to 113.98 TFLOPS; and an RTX 3080 10GB reached 110-113 TFLOPS in FP16/BF16 tests. In real AI workloads, ComfyUI tests on an RTX 3090 Ti showed around 56-58% higher BF16 throughput in several image-generation models, while INT8 gains were minimal. LLM inference gains vary, with one RTX 3090 report showing 33% higher vLLM prompt processing and throughput but little change in token generation.

GPU Unlock is experimental. It does not increase memory bandwidth or CUDA core count, and results are not universal; one RTX 3080 Ti 20GB user reported no BF16 uplift. An experimental Windows version uses a UEFI-based approach and showed about 82.5% BF16 improvement on an RTX 3090, but it is limited to specific hardware.