Advice sought on hardware upgrade

I am currently working on 2 C870s, but my supervisor wants more and faster processors and has asked me for advice on what to upgrade to. Here is his wish;

“can you look at in what way we could upgrade what kit you’ve got to get much faster performance. I am quite keen though that when we do upgrade it is not obsolete within 3 months say. Ideally, with OpenCL and Snow Leopard coming soon it would be good to see if we can get a kit upgrade that is compatible with that, much more powerful, and will be current for say a couple of years. Generally, we need the highest number crunching capability possible with an ability to post process the results with good graphics.”

I would welcome any suggestions, particularly from Nvidia employees. I think my supervisor is more concerned with not buying hardware that will be obsolete by the end of this year, and there is no requirement for graphics output capability for simulation-time visualization but post-simulation visualization would be useful. Not sure about the upper limit on the number of processors but he is looking at buying a rack of several processors.

Well, given that NVIDIA doesn’t officially preannounce products very often, it will require some psychic powers to ensure that your device won’t be “obsolete” in 3 months. :)

Picking the best hardware is now a bit challenging, as that is a multifaceted question. If you are talking about inserting cards into an existing workstation:

  • Telsa C1060 has the most memory of any single CUDA device, otherwise very similar to GTX 280/285.
  • GTX 285 has the same number of stream processors as the C1060. It is the fastest single device in terms of FLOPS and memory bandwidth.
  • GTX 295 is two slightly slower devices packed into a double-slot card. Would let you get more CUDA devices into your existing system if that is helpful to you.

I think those are pretty much now your options. Each is the best in its own way, so you’ll have to pick what is most important to you (RAM, clock speed, or # of CUDA devices).

Also, note that all of those cards use more power than the C870, so check your power supply and the required power plugs in each case.

If you are going to rackmount, then there is one solution: the Tesla S1070 is a 1U rackmount enclosure with 4 C1060 in it.

Since you mentioned Snow Leopard, does that imply that you’re using Macs?