NVIDIA CMP 170HX Mining Cards Are Being Turned Into AI Hardware

A new community tool is letting owners of NVIDIA's CMP 170HX mining cards experiment with AI workloads, but the unlock only works reliably on some boards and comes with a few hard limits. The tool was published on GitHub by amoghmunikote, and it is said to restore full SM compute throughput and unlocked HBM2e memory geometry.

The CMP 170HX first launched in 2021 as a cut-down A100-based mining card, with 8 GB and 10 GB VRAM versions. According to the reports, the 8 GB model can be unlocked to 64 GB, while the 10 GB model can reach 80 GB. That said, the result is not guaranteed. The card is based on a lower bin of the original A100 Ampere GPU, and some units use defective or subpar HBM2e dies, which means not every card will expose the full memory stack after unlocking.

Reported stable results are often lower than the maximum, with 32 GB to 40 GB described as more realistic. The hardware also has its own limits: the chip supports only INT8 data formats, without FP8, FP6, or FP4, and its performance is rated at 48 TOPS.

There is another catch on the connectivity side. The CMP 170HX PCIe interface is limited to Gen2 x4 unless the board is physically modified, and the unlock process itself reportedly requires Linux, root access, nvidia-open 610.43.0x, and disabled Secure Boot.

The unlock tool has also pushed secondary-market prices sharply higher. Cards that once sold for about $100-$200 are now reportedly listed on eBay for $1,200-$2,000, and enthusiasts are already using CMP 170HX boards to run AI models despite the restrictions. Cheap mining cards have become expensive hardware lottery tickets almost overnight.

Would you gamble on a CMP 170HX if the unlocked memory still depended on the chip you received?