Modders unlock higher compute throughput on GeForce GPUs

Modders have unlocked higher instruction issuing rates on NVIDIA GeForce RTX 30-series and RTX 40-series graphics cards, which share silicon with professional GPUs but face firmware restrictions on compute performance. The changes accelerate matrix operations used by neural networks and large language models.

For Ampere cards, a researcher developed an unlock that doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations. The GeForce RTX 3090 Ti achieves up to 80% faster matrix multiplication in BF16. Supported chips include GA102, GA103, GA104, and GA106.

A separate Ada Lovelace project demonstrated a 2× speedup in FP16 and BF16 tensor throughput with FP32 accumulation. In testing, the GeForce RTX 4090 delivered 165.6 TeraFLOPS before the unlock and 331.6 TeraFLOPS afterward. That level of computational performance is typically associated with professional hardware such as the NVIDIA RTX 6000 ‘Ada Lovelace.’ The unlock supports AD102, AD103, and AD106, targeting compute-sensitive matrix multiplication and accumulation rather than demonstrating a general increase across all workloads.