Modders have developed a firmware modification that removes performance restrictions on NVIDIA GeForce RTX 30 and RTX 40 series graphics cards. These consumer GPUs share the same silicon dies as professional-grade hardware, but their compute potential is limited by the manufacturer's firmware stack.
For the Ampere-based RTX 30 series, the modification targets GA102, GA103, GA104, and GA106 chips. It doubles the instruction issuing speed for FP/BF16 FMA into FP32 operations. Testing on a GeForce RTX 3090 Ti showed an 80 percent increase in matrix multiplication speed for BF16 operations.
The Ada Lovelace-based RTX 40 series, specifically models using AD102, AD103, and AD106 silicon, received a similar unlock. This modification doubles tensor throughput for FP16 and BF16 operations with FP32 accumulation.
During tests with the GeForce RTX 4090, the modification increased tensor throughput from 165.6 TeraFLOPS to 331.6 TeraFLOPS. These performance levels are typically restricted to professional-grade hardware like the NVIDIA RTX 6000 Ada Lovelace.
