NVIDIA RTX GPUs Were Running AI at Half Speed: Free Tool Fixes Firmware Cap on Windows - Tech Times
A Windows port of the gpu-unlock tool has been released, enabling consumer NVIDIA RTX GPUs to achieve up to double their AI compute performance by removing a firmware-level cap.
The Windows port of the gpu-unlock tool reached its third public release on October 24, bringing AI compute gains to Windows users that were previously only available to Linux users. The technique removes a firmware-level instruction-rate cap that NVIDIA applies to consumer GeForce GPUs, cutting their Tensor Core AI math throughput to roughly half the rate delivered by professional workstation cards built from the same silicon.
Community testing documents significant performance increases: on an RTX 4090, FP16 throughput doubled from 165.6 to 331.6 TFLOPS, while an RTX 3090 saw a 78.5 percent increase from 70.55 to 125.91 TFLOPS. The project has accumulated nearly 6,800 GitHub stars and 709 forks in its first eight days.
NVIDIA's Dual Pricing Strategy
The performance gap gpu-unlock closes is not accidental. NVIDIA's consumer Ampere dies contain the same Tensor Core hardware as professional Quadro and RTX Pro cards, but the firmware limits the rate at which FP16 and BF16 instructions are issued on consumer cards to roughly half that of professional products. This strategy extends back to at least the Pascal generation for FP64 compute.
The RTX Pro 6000 Blackwell workstation GPU is priced at $13,250 or higher, while the consumer RTX 4090 sells for roughly $1,000 to $1,600. The AI compute boom has made the FP16 and BF16 performance gap newly visible, particularly for researchers and hobbyists running local language models and image-generation workloads.
How the Unlock Works
The tool modifies two hardware registers (SS0 and SS1) that function as speed selectors for the Tensor Core instruction pipeline. On Linux, a kernel module writes to these registers before NVIDIA's proprietary driver loads. The Windows port, developed by impacifist, uses a UEFI application that runs in the pre-boot firmware environment, applying the same register modifications before Windows starts.
The register writes that gpu-unlock performs are narrow and targeted. Two hardware registers on NVIDIA consumer Ampere GPUs — designated SS0 and SS1 in the gpu-unlock project's GitHub README — function as speed selectors for the Tensor Core instruction issue pipeline.
Version 0.3.1, released October 24, is source-only and requires manual navigation to the EFI application from a boot menu at each system start. Automatic chainloading is not yet implemented.
Performance Gains and Limitations
Community testing shows consistent, substantial gains across the RTX 30 series in AI compute workloads. The RTX 3090 Ti climbed from 87.5 to 157.4 TFLOPS in BF16 matrix multiply (79.9% gain), while the standard RTX 3090 improved from 77.5 to 129.6 TFLOPS (67.2% improvement). Gains range from roughly 25 to 81 percent depending on card model and power configuration.
The benefits apply specifically to compute-bound workloads where Tensor Cores are the bottleneck, including large language model prompt processing, Stable Diffusion inference, and direct GEMM workloads. Memory-bound tasks like LLM token generation during decode and gaming show little or no improvement. One user reported a 16% improvement in prompt processing speed but found token generation largely unchanged.
Community testing suggests the same register mechanism applies to NVIDIA's current RTX 40 series, with the RTX 4090 showing FP16 throughput doubling from 165.6 to 331.6 TFLOPS. This performance places an unlocked RTX 4090 in mid-tier data-center inference territory, potentially saving users thousands in cloud computing costs for sustained AI workloads.
Related
