Modders Unlock Up to 2x Compute on GeForce RTX Cards
Consumer GeForce RTX GPUs share silicon with NVIDIA's professional GPUs, but their compute potential is limited by the firmware stack. Modders are now working around that limit. A researcher's unlock for RTX 30-series cards based on Ampere (GA102, GA103, GA104 and GA106) reportedly doubles the instruction issuing rate of FP/BF16 FMA into FP32 operations, with up to 80% faster BF16 matrix multiplication on the RTX 3090 Ti. For RTX 40-series Ada Lovelace cards, another project reported a 2x speedup in FP16 and BF16 tensor throughput with FP32 accumulation, lifting the RTX 4090 from 165.6 to 331.6 TeraFLOPS in testing. Why it matters: matrix multiplication underpins nearly all neural networks, so gains like these touch AI development, research and affordable local compute. It also raises questions about vendor firmware limits and product segmentation. Engineers and buyers should watch how NVIDIA responds. What's your take on unlocking hardware you already own?
Key points
- What happened Modders have unlocked compute performance on GeForce RTX 30-series and RTX 40-series GPUs that were previously restricted by NVIDIA's firmware stack.
- The key detail A RTX 30 unlock reportedly doubles FP/BF16 FMA to FP32 issue speed, giving up to 80% faster BF16 matrix multiplication on the RTX 3090 Ti.
- Why it matters Matrix multiplication is crucial for nearly all neural networks. For RTX 40, a 2x tensor throughput gain took the RTX 4090 from 165.6 to 331.6 TeraFLOPS in testing.