Nvidia RTX GPUs Hacked to Remove Half-Speed AI Throttle, Doubles RTX 4090 Tensor Performance

Modders Unlock Hidden AI Performance in Consumer Nvidia RTX GPUs

Modders have successfully exploited a firmware vulnerability to remove half-speed AI throttle imposed on Nvidia GeForce RTX 20, 30 and 40 series GPUs, effectively doubling peak Tensor Core throughput on some models1

1

2

2

. The firmware exploit targets artificial compute limits that Nvidia placed on consumer GPUs to differentiate them from professional RTX cards. Three separate open-source projects now offer generation-specific unlocks, with gpu-unlock handling Ampere cards, gpu-unlock-ada covering Ada Lovelace models, and turing-unlock addressing the RTX 20 series1

1

.

RTX 4090 Achieves 331.6 TFLOPS After Unlock

The unlock delivers dramatic improvements in peak tensor rates across all three generations. The RTX 4090 jumped from 165.6 to 331.6 TFLOPS on FP16 and from 165.5 to 330.6 TFLOPS on BF16, effectively doubling its theoretical AI performance1

1

2

2

. For the RTX 3090 Ti running at its 450W limit, matrix math improvements reached 35% for FP16, 80% for BF16, and an impressive 94% for TF32 operations2

2

. The RTX 2080 Ti saw FP16 gains ranging from 36% to 64% depending on matrix size and power limits1

1

. These figures bring consumer Nvidia RTX GPUs significantly closer to professional card performance for AI workloads.

Real-World AI Workloads Show More Modest Gains

While peak tensor rates double, real-world AI performance improvements vary considerably based on workload type and power constraints. Matrix-multiply benchmarks using cuBLAS GEMM showed the RTX 4090 achieving only 12-34% improvements in median performance, though an RTX 4070 Ti on Windows demonstrated larger gains of 80.1% and 76.4%1

1

. For practical AI applications, an RTX 3090 Ti running ComfyUI saw 56% to 59% higher throughput in BF16 image generation models2

2

. LLM prefill operations improved by 14-36% across Ampere cards, while prompt processing showed similar gains1

1

2

2

. However, chatbot token generation speeds remained largely unchanged because that performance depends primarily on memory bandwidth rather than compute capability1

1

.

Technical Details Behind the SEC2 Booter Exploit

The firmware exploit works by targeting a stack buffer overflow vulnerability in Nvidia's signed SEC2 Booter firmware, which operates in heavy secure mode1

1

. The vulnerability was discovered after Nvidia published a debug Booter encrypted with a known public test key. The exploit uses a single 4-byte write with the V67 payload to open the privilege mask, allowing the host system to modify two speed selector registers that control FP32-accumulate math rates1

1

. The module resets the card before Nvidia's driver binds it, making the changes temporary and reversible when the GPU loses power2

2

. No hardware modifications or VBIOS flashing are required, though most implementations target Linux with experimental Windows builds pursuing different approaches1

1

.

Power Limits and Hardware Compatibility Concerns

Power limits significantly impact the effectiveness of unlocking compute limits on consumer GPUs. User reports show that RTX 3090 cards capped at 225W achieved only 25% BF16 improvements, while cards running at higher power limits saw gains of 67%1

1

. The gpu-unlock project covers GA102, GA103, GA104, and GA106 series IDs for Ampere, while gpu-unlock-ada experimentally supports the RTX 4090, 4090 D, 4070 Ti Super, and 4060 Ti 16GB1

1

. Results vary significantly across different card models, with at least one RTX 3080 Ti 20GB showing no BF16 gain2

2

. Both project documentation warns that modifications may damage hardware, hang GPUs, prevent boot, or corrupt data1

1

. Any GPU modification voids warranties, and newer Hopper and Blackwell architectures remain out of scope for this exploit approach1

1

2

2

.