AI · November 11, 2026 · Low impact
Builder unlocks a CMP 170HX mining card into a working A100-class accelerator
A builder named Kate used voltage fault injection to bypass the fuse lock on a CMP 170HX mining card, then enabled its working streaming multiprocessors, from 108 on a retail A100 up to 120 to 126. Her measured FP16 throughput was 85.9 TFLOP/s versus 72.5 on a retail A100, about 1.18 times faster, with the full 80GB of memory usable for inference. NVLink and PCIe Gen4 also come back. The process needed board rewiring, a custom STM32-based glitcher and between 1 and 48 hours of glitching per card, and one of her ten cards only had 64GB intact.
It shows that cheap retired hardware can be turned into high-memory inference GPUs, though the process is far too difficult for most people to attempt.