A series of newly developed firmware exploits has bypassed Nvidia’s imposed half-speed AI performance limits on RTX 20, 30, and 40-series graphics cards. The modifications unlock native matrix-math throughput across multiple generations, enabling significant performance gains for tensor-heavy workloads such as image generation and prompt processing, though real-world application speeds and chatbot token generation remain constrained by memory bandwidth.
For all impacted GeForce generations, Nvidia’s published specifications dictate that FP32-accumulate math runs at half each card’s FP16-accumulate rate, a restriction that historically did not apply to Turing and Ampere professional-grade cards. Three independent projects now offer generation-specific unlocks using unofficial Linux modules, alongside experimental Windows builds that take a different operational approach.
The compatibility spans several generations of hardware. The gpu-unlock project targets Ampere architecture, covering series IDs including GA102, GA103, GA104, and GA106, with reported results on the RTX 3090 Ti, RTX 3090, RTX 3080 10GB, RTX 3080 20GB memory-modded cards, RTX 3070, and RTX 3060 12GB. Meanwhile, an Ada Lovelace fork designated gpu-unlock-ada supports the RTX 4090 alongside experimental builds for the RTX 4090 D, RTX 4070 Ti Super, and the RTX 4060 Ti 16GB. Additionally, the turing-unlock project—adapted from the Ampere tool—operates with memory-modded 22GB RTX 2080 Ti cards, while a separate project named KohakUnlock also targets the RTX 2080 Ti.
Performance increases vary significantly depending on card architecture, precision, and power limits. The Ada fork developer’s RTX 4090 demonstrated a peak tensor rate jump from 165.6 to 331.6 TFLOPS on FP16 and from 165.5 to 330.6 TFLOPS on BF16. Independent benchmarks show cuBLAS GEMM medians improving by 12% to 34% and up to 20% on Linux at 450 W, while a separate Windows port on an MSI RTX 4070 Ti recorded larger matrix-multiply gains of 80.1% for FP16 and 76.4% for BF16.
Ampere cards exhibit similar variance tied closely to power thresholds. A tested RTX 3090 Ti using a modified Nvidia driver at 450 W scored gains of 35% for FP16, 80% for BF16, and 94% for TF32. User reports compiled by project maintainers indicate BF16 matrix-multiply improvements ranging from 25% to 67% for the RTX 3090, 70% to 73% for the RTX 3080 10GB, 56% to 59% for the 20GB memory-modded version, 73% to 81% for the RTX 3070, and approximately 80% for the RTX 3060 12GB. The lowest figure of 25% was observed on cards capped at 225 W, underscoring the heavy influence of power limits on the unlocked performance. For Ampere hardware, large language model prefill improved by 14% to 36%, and diffusion workloads alongside FlashAttention accelerated, while LLM decode speeds and INT8 operations remained largely unchanged.
Memory-modded 22GB RTX 2080 Ti cards running Turing architecture also showed measurable gains. A Dell OEM card limited to 250 W achieved FP16 matrix-multiply improvements between 36% and 64% depending on workload size, while a 260 W variant yielded a 59% increase. BF16 remains unaffected on Turing cards due to a lack of native BF16 hardware support. KohakUnlock reported an 85% improvement on an instruction rate test using a Leadtek 2080 Ti 22GB, with a best matrix-multiply median of 53% compared against a locked card of the same model. Project maintainers emphasize that the exploit does not translate to a double-speed increase in standard applications or video games.
The underlying mechanism exploits a stack buffer overflow in Nvidia’s signed SEC2 booter firmware, a vulnerability detailed in a June research paper by Jon Pry titled “A Canary in the Crypto Mine: Defeating Stack Protection in a GPU Secure Coprocessor.” The exploit allows code execution in heavy secure mode after uncovering debug booter binaries encrypted with a known public test key. By executing a single 4-byte write using V67 payload constants derived from previous mining-card unlock tools, the privilege mask opens, enabling the host driver to configure the full rate before the standard driver binds to the card.
Developers of the open-source projects caution that the modifications carry inherent operational risks. Documentation for turing-unlock warns that the mod may damage hardware, while KohakUnlock notes that the firmware alterations can cause system hangs, prevent booting, corrupt data, or physically damage the GPU. Furthermore, newer architectures such as Hopper and Blackwell remain completely out of scope for the current methodology, and numerous board variants across supported generations remain untested or experimental.
