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% - Tom's Hardware
Modders have discovered a firmware exploit that removes Nvidia's imposed half-speed AI performance limits on RTX 20, 30, and 40-series graphics cards.
The exploit, reported by VideoCardz, restores native matrix-math throughput to affected GeForce cards, potentially speeding up AI tasks like prompt processing and image generation. However, chatbot token speeds remain largely unchanged as they depend on memory bandwidth rather than computational power.
Performance Impact
While the RTX 4090 can theoretically double its peak tensor rate, real-world performance gains are more modest. Nvidia had implemented the limitation across its consumer GPU generations, with FP32-accumulate math capped at half the FP16-accumulate rate, a restriction that doesn't apply to its professional Turing and Ampere cards.
Related
