Modders have released a firmware exploit that removes Nvidia’s built‑in AI throttle on RTX 20, 30 and 40 series cards, VideoCardz reports. The unlock lets the RTX 4090 reach twice its advertised peak tensor throughput, and the RTX 3090 Ti see up to an 80% increase in matrix‑math speed. Nvidia’s spec caps FP32‑accumulate math at half the FP16 rate on these consumer GPUs, a limit not present on its professional Turing and Ampere cards. Real‑world tasks see smaller gains, and chatbot token rates stay similar because they are limited by memory bandwidth.
The change is a firmware patch that owners can apply themselves; Nvidia has not issued an official response.
Why it matters
GPU owners can get noticeably faster AI inference on supported cards without buying new hardware.
