An open-source Linux kernel module published on GitHub by developer Zagitta allows owners of older NVIDIA GeForce graphics cards to bypass software-level throughput restrictions on Tensor Core math, lifting half-precision compute performance by up to 80% without requiring a VBIOS flash or modified driver binaries.
How the Open-Source GPU Unlock Mod Bypasses NVIDIA Restrictions
The mechanism behind the utility takes a remarkably clever approach. Rather than patching compiled graphics drivers after boot, the tool loads a custom kernel module via the initial RAM filesystem (initramfs
) before the proprietary NVIDIA driver initializes. According to technical breakdowns, this early-boot module changes low-level hardware speed-selection registers. The driver binds normally afterward, never detecting a modified binary.
This architecture builds upon earlier research originally designed to unlock NVIDIA’s CMP cryptocurrency mining GPUs. Because the modifications reside entirely in volatile configuration registers rather than permanent flash memory, the changes are completely temporary and vanish the moment the graphics card loses power.
Benchmark Gains Across RTX 20, RTX 30, and RTX 40 Architectures
While the project initially focused on Ampere-generation cards, community testing and subsequent developer contributions have expanded the technique across multiple generations. Benchmark reports published across hardware outlets demonstrate substantial leaps in dense matrix multiplication (GEMM) throughput involving half-precision math with FP32 accumulation. The tool targets support for GA102, GA103, GA104, and GA106 dies covering most cards in the RTX 30 series, alongside older GTX 900, GTX 1000, and RTX 20 series parts, while the RTX 40 and RTX 50 cards are not supported by this specific repository.
Performance results vary depending on power limits and hardware configurations. For example, a second RTX 3090 report run at a 280 W limit showed smaller gains of 43% in BF16, 38% in FP16, and 47% in TF32. Meanwhile, a compatibility report documented an RTX 3080 10GB moving from roughly 65 TFLOPS to 110 to 113 TFLOPS in FP16 and BF16 while staying inside its stock 320W envelope. With the GPU power limit configured to 315W, an RTX 3080 Ti operator documented that BF16 throughput scaled from 71.59 TFLOPS up to 113.98 TFLOPS, while TF32 capabilities climbed from 35.65 TFLOPS to 65.28 TFLOPS. An RTX 3060 12GB at a 170W power limit showed gains of 78% in FP16, 84% in BF16, and 80% in TF32. For older Turing cards, KohakuBlueleaf demonstrated an RTX 2080 Ti reaching 111 TFLOPS, up from 59.9 TFLOPS.
| GPU Model | Compute Precision | Stock Performance | Unlocked Performance | Measured Gain |
|---|---|---|---|---|
| GeForce RTX 3090 Ti (450W) | BF16 GEMM | 87.50 TFLOPS | 157.41 TFLOPS | +80% |
| GeForce RTX 4090 (Ada) | FP16 Tensor Core | 165.6 TFLOPS | 331.6 TFLOPS | 2x Throughput |
Hardware Variations Limit Unlock Success as Cloud Rates Rise
Despite impressive benchmark sheets, the unlock is far from a universal fix. Community logs show that hardware variations matter significantly. At least one compatibility report highlighted by XenoSpectrum revealed that an OEM-spec RTX 3080 Ti 20GB with a GA102-250 die successfully accepted the register adjustments, yet registered zero improvement in actual BF16 throughput.
The timing of the software release has amplified its appeal within the enthusiast community. As noted, commercial cloud providers have steadily raised hourly rates for datacenter hardware like the H100, prompting home lab operators and hobbyists to reconsider keeping older 24GB consumer flagships like the RTX 3090. VideoCardz points out that standard gaming performance remains completely unaffected by the tool, and token generation speeds during LLM text output stay bottlenecked by physical VRAM memory bandwidth rather than core compute capacity.
Warranty Risks and Future Outlook for Unlocked Silicon
Running an experimental pre-kernel module on primary hardware carries inherent operational risks, and hardware analysts emphasize that altering low-level registers will void the manufacturer warranty.
The appeal is straightforward for a specific audience: owners of aging RTX 30 hardware running local AI inference, image generation and other compute-heavy jobs.
Whether NVIDIA implements stricter validation checks in subsequent driver releases to block the initialization workaround remains the primary open question facing the open-source project.