Mods unlock up to 2× compute performance on GeForce RTX 20, 30 and 40 cards

Modders have unlocked compute capabilities on GeForce RTX 30 and RTX 40 cards that were restricted by the firmware stack, the report says. On RTX 30 cards using GA102, GA103, GA104 and GA106 dies, one modification doubles the instruction issue rate for certain FP/BF16 FMA operations into FP32. A GeForce RTX 3090 Ti consequently delivers up to 80% faster BF16 matrix multiplication.

A separate RTX 40 project reports a 2× increase in FP16 and BF16 tensor throughput with FP32 accumulation. In a test on the GeForce RTX 4090, FP16 throughput rose from 165.6 to 331.6 teraflops. The unlock is listed as supporting AD102, AD103 and AD106 dies. An update to the source text also says the same technique unlocks GeForce RTX 20-series “Turing” cards, with a reported 2× performance increase.

The two modifications pass NVIDIA driver checks and run with stock drivers; CUDA benchmarks also run normally. They must be enabled at boot, before the NVIDIA driver binds the card to the kernel, and modifying a GPU voids its warranty. The report says tested GA102 and AD102 boards used the same 450 W board power before and after unlocking.