Modders bypass NVIDIA’s AI performance limits

Modders have discovered a way to bypass a restriction that limits the computational power of certain GeForce RTX graphics cards when used in artificial intelligence applications. This method allows users to increase the performance of cards like the RTX 4090, though the actual gains vary depending on the specific task being performed.

NVIDIA’s FP32 limitation on Tensor Cores

NVIDIA imposes a limitation on Tensor Cores, specialized units used for complex calculations in AI, when performing operations with FP32 accumulation. This method allows for higher precision but results in a halved throughput compared to FP16 accumulation. This restriction applies to multiple generations of RTX cards, from the RTX 20 series to the RTX 40 series.

Exploiting a firmware flaw to unlock performance

Modders have found a way to bypass this limitation by exploiting a firmware flaw. This allows users to change two normally protected parameters, restoring maximum computational throughput without physically altering the hardware. This method is not available on professional Turing and Ampere cards, which are not subject to the same restriction.

Performance gains vary by application and card

On an RTX 4090, the theoretical tensor throughput increases significantly, but real-world benchmarks show gains of 12 to 34%. Other cards, such as the RTX 3090 Ti, show larger improvements in certain operations. However, the benefits depend heavily on the specific tasks being performed, and not all applications see the same level of improvement.

Unlocking carries risks and limitations

Projects like gpu-unlock and turing-unlock remain experimental and support only a limited range of cards. The RTX 50 series is not affected by this manipulation. Exploiting the firmware flaw can lead to system crashes, GPU failures, data corruption, or even hardware damage. While the discovery is exciting for AI enthusiasts, it is not recommended for casual users.