Nvidia RTX 20, 30, and 40-series GPUs hacked to remove a half-speed AI throttle (www.tomshardware.com)

Modders have successfully exploited a firmware vulnerability to eliminate the half-speed AI performance throttling imposed by Nvidia on its RTX 20, 30, and 40-series GPUs. This significant breakthrough allows users, particularly those with RTX cards, to unlock enhanced performance, effectively doubling the peak tensor rate for RTX 4090 and improving matrix math performance by up to 80% for RTX 3090 Ti. However, real-world workloads may not achieve the same level of improvement due to factors like memory bandwidth limitations. Three distinct projects—gpu-unlock, gpu-unlock-ada, and turing-unlock—provide unofficial modules for these adjustments, allowing users to increase their GPUs' matrix-math throughput, thereby speeding up AI prompt processing and image generation. The exploit bypasses security restrictions by manipulating registers that Nvidia's driver typically protects, making this a noteworthy development for the AI/ML community. While the performance gains are commendable, users are cautioned about potential risks to hardware integrity, as the customizations could lead to damage or system instability.