Modders have developed a firmware modification that removes performance restrictions on NVIDIA GeForce RTX 30 and 40 series graphics cards. These consumer GPUs share the same silicon as professional-grade models, but their compute potential was previously limited by the manufacturer's firmware stack.
For the Ampere-based RTX 30 series, the modification increases the instruction issuing speed for FP/BF16 FMA into FP32 operations. This update applies to cards using GA102, GA103, GA104, and GA106 chips. Testing shows that a GeForce RTX 3090 Ti achieves up to 80 percent faster matrix multiplication in BF16.
The Ada Lovelace-based RTX 40 series received a separate project that doubles FP16 and BF16 tensor throughput with FP32 accumulation. The unlock covers AD102, AD103, and AD106 silicon. For instance, the GeForce RTX 4090 increased its performance from 165.6 TeraFLOPS to 331.6 TeraFLOPS.
