cmpunlocker Turns NVIDIA's Dead Mining GPU Into a 64GB Powerhouse

An open-source patch restores 64GB of HBM2e memory and full compute on NVIDIA's crippled CMP 170HX mining card, transforming a $200 crypto relic into a viable AI accelerator.

AlphaSignalAlphaSignal
·
·
cmpunlocker Turns NVIDIA's Dead Mining GPU Into a 64GB PowerhousePRO
AuthorBy AlphaSignal Newsroom
Read2 min
TopicGpus · Open Source
SubtopicAccelerators · Kernels · Edge Deployment
  • cmpunlocker restores 64GB HBM2e and full compute on NVIDIA CMP 170HX mining cards via patched kernel modules.
  • 8GB SKUs unlock to 64GB, 10GB SKUs to 40GB, all through a single shell script and cold reboot.
  • Requires Linux x86-64, root, nvidia-open 610.43.0x driver, and Secure Boot disabled.
  • PCIe capped at Gen2 x4 in software; Gen3/Gen4 blocked by OTP fuse, x16 needs soldering 24 capacitors.
  • No FP8/FP4, ECC unresolved, 80GB unstable, some cards may have genuinely defective memory stacks.
  • Prices jumped from $100-$200 to $1,200-$2,000 on eBay within weeks of release.

A GPL-licensed tool called cmpunlocker is doing something NVIDIA very explicitly did not want anyone doing: turning its dead-end crypto mining GPU back into a general-purpose compute card with a large pool of HBM2e memory. The project targets the CMP 170HX, a card built on the same GA100 silicon as the A100 but firmware-locked into near-uselessness for anything except Ethereum mining.

The README pitch is blunt: a tool to unlobotomize your NVIDIA card. The numbers back it up.

From 8GB to 64GB with a shell script

The tool restores full SM compute throughput on the GA100-based CMP 170HX and unlocks the HBM2e memory geometry that NVIDIA restricted through firmware and OTP configuration. Running sudo ./install.sh, picking a profile, and doing a cold reboot is all it takes:

sudo ./install.sh --profile=8gb    # 8GB card → 64GB unlock
sudo ./install.sh --profile=10gb   # 10GB card → 40GB unlock

After the reboot, nvidia-smi reports the card as a standard CMP 170HX with 65536 MiB of memory and CUDA 13.0 working normally. Capacity jumps by up to 8x on the 8GB SKU and 4x on the 10GB variant, all through software.

Why the card was crippled in the first place

NVIDIA carves the CMP 170HX from binned GA100 silicon that did not meet the stringent quality requirements for the enterprise-grade A100, and officially sold it in 8GB and 10GB variants. The extra memory was disabled through firmware, along with nerfed compute performance and throttled PCIe interface speeds.

Pro article

This story is for Pro members

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

Comments

avatar