معلومات أساسية
Modders have developed techniques to bypass firmware restrictions on NVIDIA GeForce RTX 20, 30, and 40-series graphics cards, effectively doubling compute performance. These modifications target the instruction issuing rates on consumer GPUs, which share silicon with professional-grade hardware.
The Ampere-based RTX 30-series cards, specifically those using GA102, GA103, GA104, and GA106 dies, see significant gains in matrix multiplication. For instance, the RTX 3090 Ti achieves up to 80% faster performance in BF16 operations, which are essential for neural network tasks.
For the Ada Lovelace RTX 40-series, including AD102, AD103, and AD106 chips, the unlock provides a 2x speedup in FP16 and BF16 tensor throughput. Testing on an RTX 4090 showed an increase from 165.6 TeraFLOPS to 331.6 TeraFLOPS, matching performance levels typically reserved for professional cards like the RTX 6000.
These modifications maintain compatibility with stock NVIDIA drivers and pass driver checks. While the increased compute power benefits tasks like local large language models, users must enable the mods at boot. Additionally, these changes void the GPU warranty.
