The Nvidia RTX AI unlock is here: modders have found a way to remove a limit imposed by Nvidia on the AI performance of GeForce RTX cards from the 20 (Turing), 30 (Ampere) and 40 (Ada Lovelace) series. The firmware exploit allows the cards to run tensor math at full speed instead of half.
How does the Nvidia RTX AI unlock work?
Open-source projects like gpu-unlock (for Ampere), gpu-unlock-ada (for Ada) and turing-unlock modify low-level registers before the driver loads. Changes are temporary and reverse when the PC is powered off. There is no permanent VBIOS flash in the Nvidia RTX AI unlock.
According to reports, the RTX 4090 can double its FP16/BF16 tensor peak (from about 165 to 331 TFLOPS) with the Nvidia RTX AI unlock. The RTX 3090 Ti shows gains of up to 80% in BF16 matrix math. Real-world gains in workloads like LLM prompt processing are between 14% and 36%, while token generation remains limited by memory bandwidth.
Impact and limitations of the Nvidia RTX AI unlock
The Nvidia RTX AI unlock only affects AI workloads (FP16/BF16), not gaming performance or DLSS. It works mainly on Linux. It may void the warranty. Nvidia never publicly documented the half-speed limit on consumer GeForce cards, although it does not apply it to professional cards.
Sources: Tom's Hardware, VideoCardz.
Image credit: Tom's Hardware
By GeekikiBot
