Modders have unlocked compute performance on NVIDIA's GeForce RTX 20, RTX 30, and RTX 40-series graphics cards, boosting instruction issuing rates to match professional-grade GPU counterparts. Starting with the GeForce RTX 30-series "Ampere," a researcher developed an unlock that doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations on GA102, GA103, GA104, and GA106 cards. As a result, cards like the GeForce RTX 3090 Ti experience up to 80% faster matrix multiplication in BF16, crucial for running nearly all neural networks.
Next, the GeForce RTX 40-series "Ada Lovelace" demonstrated a massive 2× speedup in FP16 and BF16 tensor throughput with FP32 accumulation. When the GeForce RTX 4090 was tested, it achieved 165.6 TeraFLOPS, while the unlock pushed that number to 331.6 TeraFLOPS. These computational performance numbers are typically reserved for the higher-end professional segment like the NVIDIA RTX 6000 "Ada Lovelace."
It seems NVIDIA artificially limited the AD102 silicon's performance in compute-sensitive matrix multiplication and accumulation, affecting large language model performance. For this unlock, AD102, AD103, and AD106 are officially unlocked. Both unlocks pass NVIDIA's driver checks, override PCIe Function Level Reset (FLR), run stock NVIDIA drivers, and CUDA benchmarks perform normally. The same 450 W TGP board power is used before and after the unlock on both GA102 and AD102.
Source: TechPowerUp News · Summarized by HeadlinesBriefing