vGPU support for the CMP 170HX is ready for testing, per the cmpunlocker maintainer @_j3ns (November 4, 2026). It is the first time the project has asked for testers on vGPU, which lets one card back several virtual machines instead of being passed through whole to one. Nobody has posted a test result yet.
What was released
- The driver side stays on cmpunlocker's
vGPU-and-othersbranch, which is not in master. The branch has not changed since October 27, 2026. - cmptools is a new sub-repository (amoghmunikote/cmptools) for managing vGPU settings on the card. @_j3ns calls it a work in progress and asks for feedback.
- Going by its commit titles, cmptools is a set of shell tools behind one dispatcher, with vGPU, passthrough, status and link retrain tools. With no PCI address given, the vGPU tool acts on every 170HX in the host and gives each a 64 GB BAR1 instead of 4 GB.
- cmptools has no README, licence or release yet, so the commit titles are all there is to go on until those arrive.
How it got here
On October 26, 2026 @_j3ns said vGPU was close to working. A day later @arabus_74064 posted build notes for the branch, which the Tools page keeps. At that point the branch's three vGPU patches failed the project's build validator until they were declared in the constants file (Driver and unlock failures). Whether vGPU then worked was never posted.
What is not known yet
- Whether vGPU runs on the 8 GB card, the 10 GB card or both. No cards were named in the announcement.
- Which driver versions and hypervisors it has been tried on.
- How it behaves with Gen2 and with the 64 GB BAR1 inside guests. For a single VM, the Virtual machines and passthrough page already covers whole-card passthrough, including where to run the unlock.
How to help
If you run the branch with cmptools, post the card variant, driver, hypervisor, number of guests and what each guest saw. The board card for vGPU test reports is open for exactly that. The wiki records the first result as Reported and a second person's reproduction as a confirmation.