3080 x 2
or
a single 3090?
Can tensor flow aggregate vram across 2 cpus?
I’m assuming you mean gpu’s. And even if it could, you would probably need NVLink for that. But NVidia axed it on 3080… so NVidia says no… 
DMA over PCIe still may be possible with BAR promises… but real question is where you gonna get 2x3080? I mean, one is miracle as I understand
3090 of course, as if you only want to train one model at a time, if you want to train 2 models in parrallel, then 2 3080
Data crunching wise more VRAM is important if you’re dealing with large amounts of data, two 3080s won’t share memory however you could in theory like previous GTX/RTX 20-series get more value for compute dollars as long as you don’t reach the VRAM wall. It really depends upon the workflow, if you’re doing real-time crunching the faster GPU is going to make a huge difference.
In my opinion/experience crunching data sets such as financial(stock data) non-real-time memory becomes a limiting factor based on the larger data set–real-time crunching with a slight delay is workable with a mid-tier GPU. On the other hand if you’re doing data sets on fluid or air dynamics, you’ll be wanting that extra VRAM. (I’ve managed to get a reasonable amount of work done with a 1070/2070, the key point is to try to scale the type of workflow to the hardware)
This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.