1 titan RTX or 2 2080TI NVlinked?

morning,

I have the opportunity to either upgrade the GPU to a titanRTX or buy a 2nd 2080TI and nvlink together and improve the cooling system (it works out at about the same price).

The question is which is better? a single titan RTX does give the benefit of improving the speed of models by about 10-20% over a single 2080Ti, however this is alot slower than 2 2080TI’s. The next benefit is the significant increase in GPU memort (24GB vs 11GB), however can you use the NVlink and pytorch to combine the 2x2080ti into behaving as a single 22GB 2080Ti, as all the multi GPU benchmarks keep the batch sizes the same for 2x2080ti?

I keep hitting the upper limit of the CUDA memory and the only way to get over this is to keep reducing the batch size, but there will come a point where this is no longer viable.

This leads me to the next question, is the increase in speed on the titan just down to the increased memory size?

Hi,

I think the difference will greatly depend on your workload indeed.
If you need the memory, it is a simple choice as there is no way that I’m aware of to merge two cards together.
The biggest benefit I would see from having two cards though is that when running cross validation (or any parameter search). It is much simpler to run two job at the same time. And this will most likely be faster than the single card (as long as you have enough memory on each).

This leads me to the next question, is the increase in speed on the titan just down to the increased memory size?

No I think it’s unrelated. The die is just bigger with more compute cores.

Thanks, thats what i understood the situation to be.