Hidden AI Speed Limit Found in Nvidia RTX GPUs: Firmware Exploit Frees Tensor Cores - Tech Times
Three open-source Linux projects have removed a deliberate performance restriction in Nvidia's consumer RTX graphics cards, using a firmware vulnerability to unlock Tensor Core throughput previously reserved for professional products.
The modification, which has gone viral among GPU enthusiasts in the past week, delivers significant performance improvements for local AI inference workloads. It exploits a vulnerability in the GPU's secure coprocessor firmware to bypass Nvidia's artificial throttling of Tensor Core operations that multiply FP16 or BF16 numbers and accumulate results in FP32—precisely the operations dominant in large language model inference and AI image generation.
Identical Silicon, Different Performance
Consumer RTX cards share silicon dies with their professional counterparts—the RTX 3090 uses the same GA102 die as the Nvidia RTX A6000, and the RTX 4090 shares the AD102 die with the RTX 6000 Ada Generation. Yet since at least the Turing architecture launch in 2018, consumer cards have run Tensor Core operations at half the instruction-issue rate of professional cards.
This restriction is enforced not in silicon but in speed-selector registers controlled by Nvidia's signed Booter firmware, which runs on the GPU's SEC2 secure coprocessor. The standard driver initializes these registers to consumer rates at boot, and only code running in the GPU's 'heavy secure' execution mode can modify them.
Firmware Exploit Mechanism
The exploit was documented by security researcher Jon Pry in a paper published on Zenodo in June 2026. Pry discovered that Nvidia's signed SEC2 Booter firmware contains a stack buffer overflow vulnerability, first found on the CMP 170HX mining card after Nvidia published a debug version of the Booter encrypted with a known public test key.
the exploits… aren't my ideas by any means and that the unlock just flips the two bits needed.
A Linux kernel module loads during the initramfs boot phase, performs the firmware-level register write, resets the card, and then exits. When Nvidia's unmodified driver initializes, it finds a GPU already running at professional rates and operates normally—no VBIOS flash or driver modification required.
Performance Gains by GPU Generation
Three linked repositories target different architectures: gpu-unlock for RTX 30-series (Ampere), gpu-unlock-ada for RTX 40-series (Ada Lovelace), and turing-unlock for RTX 20-series (Turing). On an RTX 3090 Ti, BF16 throughput climbed from 87.5 to 157.4 TFLOPS—an 80% gain. The RTX 4090 showed FP16 instruction rates doubling from 165.6 to 331.6 TFLOPS, though real-world gains were more modest at 12-34% for FP16.
Turing hardware lacks dedicated BF16 circuitry, so gains are FP16-only, ranging from 36-64% depending on matrix size. RTX 50 (Blackwell) cards are explicitly out of scope for this approach under the current exploit path.
What Actually Gets Faster
The improvements apply only to compute-bound operations. When processing prompts—multiplying large matrices—the bottleneck is Tensor Core dispatch speed, which the unlock accelerates. Community testing reported 14-36% faster prompt processing in real LLM inference tasks. However, during token generation, the bottleneck shifts to memory bandwidth, which remains unchanged.
Workloads benefiting most include AI image generation with BF16 models, LLM fine-tuning, and long-context prompt processing. One ComfyUI pipeline showed 47-60% higher throughput on BF16-heavy workloads, while the MiniMax H3 model reported 60% higher throughput in BF16 mode and 31% in INT8 mode.
Related
