Will it be better to use one good GPU like the Titan X with 12 GB Memory or two GTX 1080 with 8 GB Memory each to compute large CNNs. Both will cost about the same.
Nvidia suggests the Titan X here, but why? What are the advantages having one GPU with much Memory than having two smaller GPUs which have more Memory when combined?
So from my personal experience is tensorflow tends to only use one GPU. That might not always the case, but for the algorithms I'm running it seems to use only one.
Personally I would go for two gtx 1080 ti's or one gtx 1080ti. The cost for the titan for one more GB or gddr and a slightly faster bus doesn't make since.
I would go for two just so you can run multiple experiments at once.
