Skip to content

Prerequisites

Shop and stage everything below before you treat the CMP 170HX as a daily driver. Missing cooling or the wrong power pinout will stop you colder than a bad driver pin.

Essential hardware

The card

NVIDIA CMP 170HX

  • Where used cards show up: surplus GPU markets, eBay, regional mining liquidations
  • Identify before paying: PCI ID 20c2 (8 GB → 64 GB unlock) or 2082 (10 GB → 40 GB unlock) on a stock driver
  • Avoid: boards with bent PCIe fingers, cracked cooler mounts, or mystery “already unlocked” claims you cannot verify
  • MSRP context: launched around 2021-09-01 near $4299; street pricing since The Merge has been far lower and volatile

Bring a known-good x86 Linux box to the meetup if you can run lspci before cash changes hands.

Host system

Minimum useful host:

  • PCIe x16 mechanical slot (stock still trains x4. Use an x16 slot so the link can go wider after the AC-coupling capacitors are fitted)
  • UEFI firmware with Above 4G Decoding
  • Ability to disable Secure Boot
  • Enough chassis airflow for a dual-slot passive brick, or room for a GPU waterblock and radiator

Workstation and used server boards are both fair game. Document your exact BIOS knobs on Host BIOS when you find them.

Power supply

12 V PSU with a real 8-pin CPU/EPS cable (or a quality adapter that lands on the card’s EPS-style plug)

Budget item Guidance
GPU TDP default 250 W
Software power ceiling ~300 W
Idle GPU ~30–40 W
System headroom Size the PSU for GPU + CPU + drives with ≥20% margin

Connector shape

Stock feed is 1× 8-pin CPU/EPS-style via adapter, matching Full Specs. Do not assume a standard PCIe GPU 8-pin pinout on the card side. Use the adapter that shipped with the card or a verified equivalent.

Power details →

Cooling

Pick one path before first CUDA burn-in:

Path What to buy Notes
Forced air High static pressure 120/140 mm fans aimed at the passive sink Works in open benches and server chassis with directed airflow
Water Bykski N-TESLA-A100-X-V2 + loop (rad, pump, fittings, coolant) Compatible family also covers A100 40GB PCIe / A30 / L40 per community validation
Avoid Bykski N-TESLA-A100-80G-X-V2, non-V2 acrylic block Wrong PCB or weaker construction

Thermal pads on empty footprints

Every unpopulated IC footprint under the waterblock pillars needs a thermal pad. Bare copper under metal pillars shorts and kills boards. This is the most common waterblock brick mode called out in the watercooling install guide.

Air Cooling → · Water Cooling →

Tools for optional mods

PCIe capacitor mod (later):

  • Hot air or fine tip iron suitable for 0402
  • Magnification
  • 24× (order ≥30) 0402 0.22 µF X7R, ≥16 V (example: Samsung CL05B224KO5NNNC)
  • Teardown tools: Phillips, metric hex keys, plastic spudger

Teardown / waterblock:

  • Phillips screwdriver
  • Metric hex set (Bykski’s included wrench is often the wrong size)
  • Plastic spudger
  • Organized tray for spring screws and washers

Deep teardown steps: Teardown and the Teardown Guide.

Essential software

Linux distribution

Any modern distro you can pin is fine if it runs nvidia-open 610.43.02 or 610.43.03 cleanly. Ubuntu LTS-class hosts are common in unlock write-ups. Snapshot the rootfs before patching modules.

Required pieces:

  • Root access
  • linux-headers-$(uname -r) / kernel-devel matching the running kernel
  • Build tools for the unlock installer
  • Python 3 (cmpunlocker build-time profile selection)
  • Network on first install

Unlock tooling

Project Role URL
amoghmunikote/cmpunlocker Canonical community unlock https://github.com/amoghmunikote/cmpunlocker
bayley/cmpunlocker Fork with additional P2P work https://github.com/bayley/cmpunlocker
Consensus-Protocol/cmp170hx Deep wiki (mechanism, procedures, frontier) https://github.com/Consensus-Protocol/cmp170hx

Driver pin

Cache the exact nvidia-open 610.43.0x packages locally. Community notes warn that NVIDIA may drop older open builds from easy mirrors.

Monitoring

Install after drivers load:

  • nvidia-smi
  • nvtop or equivalent
  • lm-sensors if your motherboard exposes useful package thermals
  • Your LLM stack’s own metrics (vLLM / DFlash / whatever you standardize on)

Thermal materials

  • Fresh paste for die / heat spreader work after teardown
  • Extra thermal pads sized for VRM / PMIC / empty footprints when watercooling
  • Isopropyl alcohol and lint-free wipes

Storage and networking

  • Fast system NVMe for model weights
  • Wired Ethernet while you debug unsigned modules and Secure Boot

Spares

  • Extra 0402 caps if you attempt the lane mod
  • Spare EPS cable
  • Known-good display GPU or BMC/IPMI path so a black CMP (no outputs) never leaves you blind

Optional but useful

Item Why
Second identical 170HX Multi-GPU / P2P experiments (fork-dependent)
Dedicated bench power meter Correlate wall draw with the 250–300 W envelope
USB microscope 0402 inspection
Chassis with 2.5–3 slot clearance Waterblock + fittings height

Before you order checklist

From existing hardware

  • Host with a free PCIe x16 slot and working Linux install path
  • Monitor/KVM that does not depend on the CMP (it has no display outs)
  • Keyboard, another machine for ISOs, USB installer stick
  • Screwdriver + spudger if you plan teardown

Plan the build

  • Airflow map or radiator placement sketched
  • PSU EPS cable confirmed
  • SKU target chosen (64 GB unlock vs 40 GB unlock)
  • Driver version 610.43.02/03 download staged
  • Unlock repo cloned on a machine that can reach GitHub

Ready to proceed?

  1. Quick Start for the ordered bring-up
  2. Introduction if you still need the SKU story
  3. Specifications when you write purchase notes or bench headers

Lab tip: Order cooling with the card. The stock heatsink is a passive server part and needs directed airflow.

References