Motherboards and device to device bandwidth speeds

ashj

Commendable
Jun 25, 2016
2
0
1,510
Hi,

I'm thinking of building a 2 or 4 GPU system to run deep learning neural networks on. Deep learning requires a great amount of device-to-device PCI express bandwidth. I've read that the device-to-device bandwidth is largely determined by the PCI express switch that is in the motherboard. If you have a motherboard with four x 16 slots and one PCI e switch connecting then all, then you each device should theoretically receive 16 PCI lanes to one another.

How do I determine which motherboards have the best PCI switches for a Deep learning machine that requires the maximum device-to-device bandwidth?
 
Solution
if you wanna do those things with multiple gpu, i suggest you getting xeon cpu with server motherboard, mainly because it pci lanes support is a lot, the reliability, ecc memory support so you get better stability, and also more features that xeon have, but for sure it wont be cheap and not as easy as getting desktop components

nvincent_08

Reputable
if you wanna do those things with multiple gpu, i suggest you getting xeon cpu with server motherboard, mainly because it pci lanes support is a lot, the reliability, ecc memory support so you get better stability, and also more features that xeon have, but for sure it wont be cheap and not as easy as getting desktop components
 
Solution