GPU not being used

Is a portable external SSD fine? Or would this slow it down?

I’m not an expert and maybe @rwightman might throw in his opinion, but I think it depends on the USB controller etc. If you have an external SSD handy, go for it and compare the results, but an internal connection will still be faster. (Not sure about eSATA in case you are using it, which should be also pretty fast)

If you’ve got a portable SSD, try it. Even with an older USB 2 interface, it’ll still beat a spinning disk on seek time.

If you’re going to buy something for this purpose, I’d try to avoid portable if you can. That being said I do use SATA SSD drives with the Startech SATA -> USB 3.0 or 3.1 converters for moving data between machines, or getting data to embedded NVIDIA Jetson boards for inference test data. It’s a pretty good setup but I’ve never tried pushing them to the level I would for training


1 Like

One other comment here, based on your network arch, it looks like the images are REALLY REALLY small. Something like 12x12 to 15x15? That’s smaller than MNIST. So, surprised they don’t fit into memory, especially if the epoch counts and batch size numbers in the snippet posted are for the full dataset?

With basic MNIST python training loops, data all in memory, you can usually only push a higher end GPU into the 30% utilization range and have to get creative to go further.