After some recent experimentation and research, I have not found explicit hardware requirements for a platform to support Compute (No Display) mode on RTX 6000 Pro Blackwell Workstation cards.
POST failures are seen on platforms that are found to be incompatible, requiring obtaining a compatible platform to reflash the cards back into display mode to allow POST.
I have seen reports of (and experienced) this in quite a few instances where end users must trial-and–error plugging the card into different machines until one is found that is compatible and does not induce POST failure in compute mode.
The general consensus found seems to be “don’t use old hardware”, but there does not seem to be a clear definition of what is “old” and what the “minimum requirements are”.
Is anyone able to point to a resource that lists what capabilities a platform must have to support Compute Mode? Even with Above 4G Decoding, ReBAR, SR-IOV, IOMMU, MMIO, etc. enabled, it seems to be a coin flip on whether or not a POST failure/incompatibility will be seen in compute mode.
Without the ability to enter compute mode, MIG shows as disabled.
Hi, I agree that it is troublesome to find explicit information on this topic…
The issues arises NOT just from one or more features to binary be just working/avail or or not, but its like this:
With the change to Virtualization ON = display OFF mode for compute comes a change in the BAR1 size:
active cooled (=workstation) RTX PRO 6000 boards come with a default (Virt OFF, display ON) of BAR1=256MB, so requesting a contiguous 256MB window of addresses from systemBIOS. This is the default for the workstation boards, the server board CAN be changed to this mode, and BAR1 size, to allow for the displays out to really drive a display from the server board….
[display mode selector tool allows for this BAR1 to be changed to 8GB, for transfer-heavy applications and workloads - IN SUPPORTING systems…]
Changing this to Virtualization ON = Display OFF changes the BAR1 size to 64GB! This is the default the server version of the RTX PRO 6000 ships with…
Now, whether ‘your’ system is able to provide such a large address range to one (or even more) GPUs is purely to the SystemBIOS developer to ‘implement’, its NOT a hard Yes/No on one of the features you mentioned above…
As a rule of thumb ANY system in our certified list of systems (which always lags behind with newly released systems) that is listed to support a RTX PRO 6000 server version will for sure manage that 64GB addresses request - though these systems are typically designed for passively cooled server cards, so not a classical workstation config :-(.
I’d expect for any workstation certified system for the active RTX PRO 6000 to ALSO support that GPU in the display OFF = 64GB BAR1 mode, but that requires another GPU next to the RTX PRO 6000, since it no longer can drive any display out now…
I know this is NOT a solution to this topic, but I hope I could at least give some explanation as to why this is not just ‘easier’…?
regards
-Frank