I've discovered a couple of Tesla V100 GPUs and I'm attempting to set up an AI rig, but I've hit a roadblock. My system can boot with a single V100 or with a V100 and a 2060 Super, but when I try to boot with both V100s installed, I end up in a power loop without reaching the POST screen.
Here's the current setup:
- Gigabyte B550 Eagle WiFi 6 motherboard
- Ryzen 3600X CPU
- Zalman ZM1250 power supply
- Various RAM sticks that are somewhat dubious in quality.
Initially, I dealt with a troubling SoDIMM in an adapter which only allowed booting at 2133MHz. I obtained two sticks of 2133MHz DDR4 RAM, but regardless of the combination, the system fails to boot with two V100s.
Interestingly, it worked at 3200MHz with no GPU but defaults to 2133MHz when a GPU is present. I've tried a different 1000W PSU and lowering the RAM speed to 1866MHz with no success. I'm baffled by this issue and would appreciate any guidance!
3 Answers
I had a similar issue when I was setting up a mining rig. The top PCIe slot always has full lanes, but the other slots are limited. Make sure the V100 is in the top slot or try switching it to one of the lower slots without other cards. If you still have problems with two V100s, it might be a limitation of the motherboard or CPU.
It's worth noting that your Ryzen 3600X might not have the best support for those enterprise GPUs. I switched to a Ryzen 5600X for similar issues and it fixed my boot problems completely. If you have access to a 3700X or 5600X, see if swapping out the CPU helps.
It sounds like you're hitting issues with the memory address space when running enterprise GPUs like the V100. These GPUs are known to reserve a lot of space, which might exceed what a consumer setup can handle. Make sure you have resizable BAR enabled, as it can help with these types of cards. You might want to look into adjusting PCIe bifurcation settings too, as that could influence how the lanes are allocated, particularly on the B550 board you're using.
Thanks for the advice! I've tried both the top and lower slots with different configurations, though it still boots with a single V100. I'll look into testing the 1x slots for the two V100s next.