An open-source project called GPU Unlock is stirring the local AI inference community by unlocking FP16 and BF16 instruction issue rates on NVIDIA's consumer-grade GeForce RTX 3000 series GPUs, elevating them to the level of NVIDIA's professional equivalents. Created by a developer known as "zagitta," the project is a Linux kernel module that operates before the proprietary NVIDIA driver loads, selectively modifying firmware speed selectors on the GPU to unlock double the speed of FP/BF16 FMA (Fused Multiply-Add) instruction emissions for FP32 operations on GA102, GA103, GA104, and GA106 chips.
The numbers are striking. On an RTX 3090 Ti, the project reports BF16 compute jumping from 87.50 TFLOPS to 157.41 TFLOPS — an 80% leap in semi-precision computation. FP16 climbs from 87.46 to 118.39 TFLOPS, and TF32 nearly doubles, from 42.58 to 82.61 TFLOPS. Similar results have been reported across other family GPUs: the RTX 3080 gains 70-73% in FP16, the RTX 3070 between 73-81%, and the budget-friendly RTX 3060 12GB approximately 80%. The project also supports the GTX 900, GTX 1000, and RTX 20 series, though it does not currently work on RTX 40 or 50.
What makes GPU Unlock particularly compelling for anyone running local language models is the direct impact on the inference pipeline. Users have reported 14-36% gains in prefill speed with vLLM — one of the most popular inference engines for LLMs — and a 25% reduction in Time-To-First-Token latency, a metric that end users feel directly when waiting for a local model's response. Decode operations, however, remain unchanged since they are constrained by memory bandwidth, which the unlock does not modify.
The technical method behind the unlock is elegant. The module combines the V67/PLM technique from the CMP mining card unlocker with UEFI-approach timing, executing the unlock sequence, writing speed selectors in the firmware, and performing a Functional Level Reset (FLR) before the unmodified NVIDIA driver assumes control of the GPU. Unlike driver modifications, this means every NVIDIA driver update requires no rework — the module loads via initramfs during boot, well before the proprietary driver is loaded. Installation consists of cloning the repository, running `make install`, and appending `gpu_unlock.targets=all` to the kernel command line parameters.
NVIDIA has historically restricted certain compute capabilities and vGPU features on consumer GeForce GPUs, even when they share the same chip architecture as the more expensive professional Quadro, RTX A-series, and RTX PRO lineups. The corporate logic is straightforward: separated market segments, priced by unlocked performance. GPU Unlock exposes that this restriction is purely software-based — the capable hardware exists, but the firmware imposes an artificial limit to maintain price differentials between products.
This raises questions that extend well beyond benchmarks. If the hardware can do the work and the limit is merely a bit in firmware, why does NVIDIA maintain this restriction? For the user running local inference — the audience of MPDev — the answer matters because it directly affects the cost-benefit calculus of used GPUs. If a used RTX 3090 Ti at a fraction of an A100's price can achieve comparable performance in dense matmuls with a kernel module, the economics of anyone training or inferring models locally shift significantly.
The project is GPL-2.0-only with source code under public scrutiny. NVIDIA has not publicly commented on the release, but the precedent exists: the community has already unlocked similar capabilities on CMP mining cards, and NVIDIA eventually adapted. What GPU Unlock demonstrates is that the boundaries between consumer and professional hardware are less rigid than marketing suggests — and that in the world of local AI inference, open-source community creativity remains the most valuable resource.
Sources: WCCFTech, AI Weekly, GitHub - Zagitta/gpu-unlock
✓ Independent sources cross-checked and verified before publishing
