Nvidia RTX 20, 30, and 40-series GPUs hacked to remove a half-speed AI throttle — firmware exploit doubles RTX 4090's peak tensor rate, lifts RTX 3090 Ti matrix math up to 80%

Why it matters

Modders have found a firmware exploit that removes Nvidia's half-speed AI performance cap on RTX 20, 30, and 40-series GeForce cards, VideoCardz reports. Nvidia's own specs show FP32-accumulate math running at half the FP16-accumulate rate on consumer cards, a restriction not applied to its professional lines. The exploit abuses a stack buffer overflow in Nvidia's signed secure-boot firmware, discovered after the company published a debug build encrypted with a known public test key, letting a small payload open the privilege mask so the host driver can set full rate. Three projects distribute unofficial Linux unlocks covering different generations, with experimental Windows builds taking another route, and coverage varies widely by card. Peak tensor rates improve dramatically on paper, but real workloads gain far less, chatbot token speeds barely change because they depend on memory bandwidth, and power limits strongly affect results. Project documentation warns the mods can hang systems, corrupt data, or damage hardware, so caution is warranted.