Hello there,
I have an ATX PSU that powers my motherboard + 1x 3080.
I also have a HP server PSU that powers an additional 7x 3070.
Both PSU are on the same circuit. They are not linked together (server psu does not have a 24 pin)
I have Restore AC Power Loss to Power ON within bios.
On power loss and restore, both PSU power on and all 8x GPU fans begin to spin.

Problem-
Only the first GPU (the one mounted on the mobo + powered by the ATX psu) is recognized after the automatic reboot.
The gpu's powered by the server PSU are not recognized even though they are "on" (fans spinning, green light on riser).
I believe this is due to the ATX power supply starting just a fraction faster than the server psu. I believe that to be the case because it works perfect as long as I manually restart after the power failure (because the PSU "is always on" at that point).
If I then manually restart the rig all GPU are then recognized immediately and all is good.
The problem only happens during the first automatic restart after a power outage.
What's the solution so that I can rely on the automatic restart instead of having to perform an additional manual restart of the mobo after each power failure?
I changed the post delay from default (3) to (10) although I believe that only takes into effect during a normal post, so therefore isn't useful during a restore from ac power loss boot.

Thanks for all help.

Is this a crypto mining rig? Are you not using a 24-pin splitter?

Oh, the second PSU doesn't have a 24-pin to begin with...

How is the primary PSU connected to the server/secondary PSU?

This post was edited by Xet on Jun 30 2021 08:13pm

I never used a server PSU with a break out board. Mainly because I don't have that many GPUs to begin with. But after reading through, it seems like these break out boards don't have the ability to detect small power cycles such as restart/reboot. I thought the Molex connection going into the primary PSU would suffice to detect the restart but after reading through their manuals and specifications, that's not the case. I don't think even if you got the break out boards with a 24-pin, a restart/reboot will work on the server PSU.

It looks like the only solution is to do a manual or at least mimic a manual power cycle. The break out board's Molex can achieve that through using another board and this second board is controlled through LAN. But this solution is no different from using a server rack PDU, other than perhaps the cost.

If you're restarting due to performance issues, then I think the best way to achieve a power cycle here is through LAN. If it's because your power utility grid is not 99% reliable, then perhaps use a backup power supply.

Yea the conclusion i've come to is either backup power supply or continue to do it via remote connection. I've settled on remote connection because a backup power supply for 5-10 minutes 3x weekly of over 3000W is prohibitively expensive.
Thanks for your input.

Don't know what kind of OS you're running but you could probably create a startup script that'll reinitialize your GPUs or do an additional reboot of your system so they get recognized. Kind of a workaround but you're probably not going to encounter this very often especially if you have a UPS.

This post was edited by TimmyTurner on Jul 2 2021 08:42am