Nvidia’s old CMP 170HX mining cards have suddenly become AI bait after a tool has tipped up which unlocks huge chunks of memory.
The CMP 170HX launched in 2021 during the crypto mining craze with a cut-down A100 Ampere setup and 8GB or 10GB VRAM options. After the mining crash, the card looked fairly useless, mainly because Nvidia had hobbled it as a mining-only bit of kit.
Now a GitHub tool called “CMPUnlocker” claims to restore full SM compute throughput and unlock HBM2e memory geometry. The tool reportedly pushes 8GB cards to 64GB and 10GB cards to 80GB, although that comes with a fat silicon lottery warning.
The reason is that the CMP 170HX uses lower-bin A100 hardware, with many HBM2e stacks disabled because the memory dies were defective or ropey. Some cards may only manage 32GB or 40GB reliably, while others can expose more memory but lose bandwidth or clock speed.
Reports suggest bandwidth sits around 700GB/s to 800GB/s per board, with Samsung 10GB models reaching 80GB but often turning flaky. Hynix-based 8GB cards appear more stable at 64GB, though stability still depends on the quality of the memory stack.
The AI catch is that the card only supports up to INT8, with no FP8, FP6 or FP4 support like newer Nvidia chips.
Prices have gone bonkers, rising from about $100 to $200 to more than $1,000, with eBay listings now sitting around $1,200 to $2,000. That makes the CMP 170HX a cheap large-memory AI card.
