Informasi penting
Modders have developed firmware exploits to remove artificial half-speed AI performance limits on Nvidia RTX 20, 30, and 40-series GPUs. By exploiting a stack buffer overflow in Nvidia’s signed SEC2 Booter firmware, these unofficial modules allow users to flip registers that restore native matrix-math throughput, which was previously restricted to pro-grade cards.
The unlock process varies by generation, utilizing Linux modules or experimental Windows builds to bypass the throttle. The gpu-unlock project covers Ampere cards, while gpu-unlock-ada targets Ada Lovelace architecture, including the RTX 4090. Turing-based cards, such as the RTX 2080 Ti, are supported via the turing-unlock and KohakUnlock projects.
Performance gains are significant but inconsistent. While peak tensor rates for the RTX 4090 can double, real-world matrix-multiply benchmarks show improvements ranging from 12% to 94% depending on the specific card, power limits, and workload. Developers warn that these modifications carry risks, including potential hardware damage, data corruption, or system instability.
Despite these gains, actual chatbot token speeds remain largely unchanged as they are constrained by memory bandwidth. Furthermore, support remains limited; Hopper and Blackwell architectures are out of scope, and many specific card variants remain unverified. Projects are currently gathering user reports to expand compatibility, though no official roadmap exists.
