Computer won't boot with 2 Tesla V100s

Hardware Hub Build a PC
,
1

OK, so I found a few Tesla V100s behind a dumpster and I’ve been trying to set up an AI rig for a few months now.

Currently, I got it to boot up with a single V100, or with a V100 and a 2060 Super, but I can’t get it to boot with 2 V100s.

I’m running:

  • Gigabyte B550 Eagle WiFi 6
  • Ryzen 3600X
  • Zalman ZM1250 PSU
  • Different flavours of shady RAM, because them’s the times

At first, I had some cursed SoDIMM in an adapter, and it took me a while to figure out that the PC would boot only if I lowered the RAM speed in the BIOS to 2133MHz. The PC would boot with the cursed RAM at 3200MHz if there was no GPU in the system.

Since then, I got 2 different sticks of 2133MHz DDR4, and with any of them, the computer only boots with a single V100, or with a V100 and a 2060 Super, but not with 2 V100s.

The PC enters a loop of power on - power off - power on… It won’t get to a POST beep of any sort. Since the symptoms are the same as when the original cursed SoDIMM wouldn’t boot, I’m thinking RAM could still be an issue. But, none of this makes any sense to me. How can the PC boot at 3200MHz with no GPU, but require 2133MHz if there is a GPU in there?

I tried a different 1000W PSU, with the cursed RAM at 3200 and a single V100, and it wouldn’t work. I don’t have access to this PSU anymore, so I can’t test all the permutations.

I also tried lowering RAM speed to 1866, no luck.

Can anyone share some wisdom please?

2

Card drawing too much power from the chipset connected PCIe slot?

1 Like
3

Hmmm is that a thing? It works when a single V100 is in either the x16 slot or one of the chipset slots. With or without the 2060 Super, in both configurations.

Btw, I just tested with a good Corsair 3200MHz RAM, same boot loop.

4

Hopping in as I just set up a NAS with this same board and had all KINDS of issues getting it to post.

What I found is that the x16 slot is the only slot on the mobo that can handle the demands of what I was trying to install. I had a P2000 GPU, a 10gbe NIC, and a HBA card and the only way it would post is if the the NIC was in the x16 slot.

The rest of the slots are all x1 and work through the chipset. Since the chipset only has like 4 PCIe 3.0 lanes, it really has issues with the hardware handshake during the post process.

Only bright side is that I only paid $200 for the mobo, a 5500 CPU, and 16gigs of memory. :wink: