GPU Recommendations
Hi everybody
I am investigating an upcoming upgrade of the computing capabilities for running DualSPHysics and looking for advise and information for what to buy.
I narrowed the search down to 3 categories:
1) Single graphic card - Buying a "gaming" computer with a GTX 1080 TI or similar
2) SLI setup - e.q. a NVIDIA DIGITS DEVBOX w. 4 graphics cards
3) Professional grad - Tesla cards
Does it make sense to offer money for Tesla card(s)? When will it make sense to invest in them? What are their performance compared to the consumer cards? I know DualSPHysics does not yet support multiple graphics cards, but as I have understood a DEVBOX could still run 4 individual simulations simultaneous?
Going further, neither less the choice of category, what graphics is the best/bang for the buck?
Hope you can help me out - in advance thank!
I am investigating an upcoming upgrade of the computing capabilities for running DualSPHysics and looking for advise and information for what to buy.
I narrowed the search down to 3 categories:
1) Single graphic card - Buying a "gaming" computer with a GTX 1080 TI or similar
2) SLI setup - e.q. a NVIDIA DIGITS DEVBOX w. 4 graphics cards
3) Professional grad - Tesla cards
Does it make sense to offer money for Tesla card(s)? When will it make sense to invest in them? What are their performance compared to the consumer cards? I know DualSPHysics does not yet support multiple graphics cards, but as I have understood a DEVBOX could still run 4 individual simulations simultaneous?
Going further, neither less the choice of category, what graphics is the best/bang for the buck?
Hope you can help me out - in advance thank!
Comments
Having run on a GTX-980 and Tesla K80, I've had my single-precision simulations run faster on the GTX-980, but I haven't tested double-precision simulations on both cards. That said, the Tesla K80 can handle much larger particle counts because it has 3x the GPU RAM. Once we get the Multi-GPU code, I look forward to being able to run large simulations on twin K80s with a combined 24 GB GPU RAM (many particles!), but we're not there yet.
In my view, it comes down to priorities. I am most productive with DualSPHysics when I am running many simulations in a row, quickly, at modest resolution, so I prefer to use the GTX-980. When I need to run a massive simulation in double precision and I know I've got my model set up exactly the way it should be, I don't mind that the Tesla takes a while to get the job done. But gaming cards are (relatively) cheap, and Teslas are expensive.
It all depends on the work you are trying to do, but if your goal is to have a good general purpose setup at a reasonable price, then here's what I would recommend:
- Buy the latest-generation NVIDIA GeForce GTX card with as much GPU RAM as you can, or:
- Buy more than one of them and use an SLI bridge, since the beta release of the multi-GPU code is on the horizon
My best advice is to first consider what your goals are with DualSPHysics and let your preferences (speed, max particle count, single or double precision, etc.) guide your decision as you factor in the cost of the cards themselves. There will always be compromises. Best of luck in your search!