Being involved in the Deep Network Training (on GTX1050 which has 640 CUDA cores and 2GB GDDR5), I'm thinking of buying a more performative video card. Are there any ways to make a rough calculation of the training performance improvement just by its specification?
Actually, that's quite a broad question. The answer depends on what kind of operations are you going to do with CPU during DL training.
Well, I'm seeking for somewhat universal solutions with the accent on distributed nets.
Well, there are many factors that affect training performance improvement, and this is not only about specifications. For example, amount of CUDA cores matters unless your training algorithm is inherently sequential (in this case library you are using will involve paralyzing). The amount of memory matters, especially if you are working with huge datasets, but taking into account convention on ML, you have to process small pieces of data at a time. That is why there's no strict answer to your question.
Just drop us an email to ...
Show more 
Log in with Google