Another question, When using both PCIe slots do all MBs switch them to 8x8 or do some keep both the slots at 16x?
That looks just fine.
Most mb split the pcie lanes to 8x8x
You can find some that don't, I think, but those are the $500+ boards.
It's also sometimes difficult to find boards with enough room to physically place 2-4 cards in simultaneously. On mine I have to remove one to remove the next, no clearance.
My bottom slot requires me to use a "mini" version of a card , otherwise it covers up the IO , cmos reset, USB 3 and other connectors.
I dunno what I'm doing
RTX 3090 : RTX 3070 : RTX 3060 : RTX: 2070 : Ghetto 1060
- Site Admin
- Posts: 622
- Joined: Fri Jul 12, 2019 12:49 am
- Location: San Francisco
- Has thanked: 3 times
- Been thanked: 159 times
The only way for a motherboard to have multiple x16 PCI-E lanes is for the CPU to have a lot of lanes. Some Threadripper and Xeon boards can support that. All consumer chips, however, have too few PCI-E lanes to support 2 x16 PCI-E ports (after some are used for the southbridge/USB ports/SATA connections and the like).
Alright.... The new setup is built. Ubuntu 20.10, Ryzen 3800x, 32gb ram, and the MPG x570 Plus mb. Interesting results on my first tests.
Tests with one gpu (did both GPU0 and GPU1) give me 20-21 EGs/sec with a batch of 7.
Tests with distributed, batch of 14 (2x7) give me 24 EGs/sec. Only a slight gain.
The singles did 4000 iterations with a batch of 7. The distributed did 2000 iterations with a batch of 14. The math still seems off. I'd expect both cards should be able to roughly double what one could put out. But it is much less than expected.
Both tests saw the same number of faces, 28,000. I watched the pci-e utilization numbers and distributed topped out at 30% of capacity. Both cards showed mid 90s% utilization.
A sizable improvement over the old MB/CPU, Just wondering what kind of numbers other people are getting. Thanks.
My feeling I'm seeing a linux driver issue or a tensorflow issue. I installed windows on the new computer and tried to run the same test.
Single gpu performance was about 10-20% worse under Windows.
I could not get distributed to work at all. Kept throwing out tensorflow illegal memory errors.
I also forgot how much I hate windows.....