I am in the process of deciding a possible update for my personal DL hardware. Having sold my 1070, I do have now just a single 1080ti.
Now, for the same amount of money (~550 eur), I could:
-
Buy another 1080ti, used, on ebay.
-
Sell my 1080ti, and buy a 2080ti.
Considerations: I work every day with 1-2 tesla V100 (32Gb version), and found that my 1080ti is not that much slower than the mighty V100… What really makes the difference is memory… Having much more freedom with batch sizes, training resolutions, and model size, that’s quite something.
Furthermore, I’m having troubles in observing those amazing speedups in fp16 on the teslas. What I observed is a speedup of some 30%, but the model usually converges a bit more slowly, so the advantages of fp16 are rather thwarted (both with plain fastai/pytorch and with nvidia apex addition).
So, I’m leaning towards a second 1080ti.
Advantages:
- I can have 22gb of “true” memory (that is, without fp16), at least for vision. Maybe, in a close future, even for other applications. Conversely, I can estimate some 19-20gb for the 2080ti in fp16, given that one can always use fp16 reliably.
- I can run more experiments at once w.r.t. a single 2080ti.
- Two 1080ti are more powerful than a single 2080ti.
Disadvantages:
- Investing more money in a relatively old hardware arch (remembed that pytorch broke support for gtx 900 series).
- Much higher power draw w.r.t. a single 2080ti.
- More difficult to resell two 1080ti within a year w.r.t. a 2080ti.
- Buying an used gpu may reserve unpleasant surprises.
- Maybe I didn’t observe substantial advantages in fp16 just because I didn’t use it the right way?
I’d be glad to have your opinions, with particular interest for those coming from rtx’s owners.
Thanks.
