new cuda system building questions

i have about $5000 to build a cuda system that will be used in fluid dynamic simulation research.

i have one major question, is the Tesla c1060 better than a GTX 295 for cuda? if it is better, does it justify the huge price difference?

The GTX 295 features 480 stream processors with 1792mb of ram (half to each gpu), a core clock of 576mhz and a memory clock of 1998mhz.

The Tesla c1060 features 240 stream processors, 4gb of ram, a core clock 1.3ghz, 800mhz memory clock.

The following two builds are what I am considering. One is based on Tesla and the other on the GTX 295. Please let me know which you think would fare better and why… Any suggestions are welcome. Thanks!

Tesla Build:

CPU: Core i7 920 ($279.99)
Motherboard: Asus P6T7 Supercomputer ($449.99)
Ram: Corsair 12GB DDR3 ($229.99)
Hard Drives: 1x Hitachi Ultrastar 15K 300gb ($319.99)
2x Seagate 1.5tb (RAID 0) ($239.98)
Video Card: GTX 285 ($369.99)
Teslas: 2x ($2599.98)
Optical Drive: Optiarch DVD Burner ($31.99)
PSU: Silverstone 1200W ($279.99)
Case: Antec 1200 ($172.99)

Total: $4,973.89

GTX 295 Build:

CPU: Core i7 920 ($279.99)
Motherboard: Asus P6T7 Supercomputer ($449.99)
Ram: Corsair 12GB DDR3 ($229.99)
Hard Drives: 2x Hitachi Ultrastar 15K 300gb ($319.99)
4x Seagate 1.5tb (RAID 0) ($239.98)
Video Card: 4 GTX 295 ($1999.96)
Optical Drive: Optiarch DVD Burner ($31.99)
PSU: Silverstone 1200W ($279.99)
Case: Antec 1200 ($172.99)

Total: $4,563.85

“Better” really depends on what algorithms you are using. If you are doing something like an n-body simulation, where the whole thing basically has to reside in device memory, then you might want to consider the Tesla (if your simulations will be large enough that they won’t fit in the 295’s memory.

However, if you’re “compute bound” (basically, very intensive computing tasks, like factoring integers or something that doesn’t require a lot of memory, go with the 4x GTX 295’s, because you’ll be getting 4 times the computing power as you would from the two Tesla system.

I don’t know much about fluid dynamics algorithms, so perhaps you could search around (or email someone at nVidia and ask) and find someone else that has done it before and see what they think.

should be

Personally, I’d only go for the Tesla series if you need the extra RAM.

N.

Like everyone has said, it depends entirely on the code you want to run. There’s no generic answer. If you need the memory, the Teslas are pretty much the only option; if you don’t (and the algorithm can scale to 8 GPUs), the GTX295s will be faster.

As a second issue, 6 hard drives (including 2 at 15k RPM) plus 8 GPUs is on the very edge if not beyond the capabilities of a 1200W PSU. A bigger PSU, or dual PSUs might be in order, which means you’d be pushing the edge of a typical 15A 120V circuit (less of an issue if you’re in a 240V part of the world). You don’t have the 4 6 pin and 4 8 pin power connectors you need for those cards, but even with adapters I’d look very very close at how many amps you’d be putting on each rail. Also, you’ll need to think about a case that has 8 slots for those 4 cards.

i am receptive to the idea of using dual power supplies, but, i am just worried about fitting them into the case.

i am thinking that the thermaltake armor plus is going to be the case that i use as it has enough expansion slots and it can fit the p6t7 motherboard.

i don’t really want to have to do any case modifications to get all the parts into the case…

it seems like i might be able to mount the second power supply in the fore section of the case where the water cooling system would go (the armor is already set up for accommodating a water cooling system(

Check out the LIAN LI PC-V2010B. It has 8 slots (check the pics, because the specs on newegg say 7) and dual PSU bays. Another option is the Silverstone 1500W, which has enough PCI-e connectors for 4 GTX295s, not that I’m personally certifying that even this 1.5kW beast is enough for that much hardware. You’ll have to pay attention to how the 12V rails are divided between all the connectors to be sure.

I’m pretty sure NVIDIA doesn’t certify power supplies for more cards than SLI supports (3 GTX285 or 2 GTX 295, IIRC), but I’d be interested to hear from tmurray or some else from inside whether they ever build boxes like this.

dual power supplies won’t be a problem in the LIAN LI PC-V2010B, as you said, and it is certainly looking like the most attractive case for the build.

however, i am not sure that it fits the CEB form factor or not. I have heard that the CEB form factor will fit in an EATX case, but, i am not sure about that. if anyone knows i would greatly appreciate clarification. i have to make sure that everything that i buy is going to work the first time!

thanks!

With a little reading, it seems that CEB is based on ATX port locations and mounting holes, with a 10.5" board length. EATX is the same thing, but with a 13" board length (and maybe another row of mounting holes?), so you should be fine. All three specs are 12" high.

If your simulation requires less than 2GB of GPU RAM, then you might want to consider another configuration with three GTX285s. Computationally, the 285 is substantially quicker than the C1060. And much less expensive.

On the P6T7 motherboard, each of the three 285s will get full PCI-E bandwidth. For most (all?) simulations of interacting components spread across multiple GPUs, three 285s will be quicker than four 295s due to gpu-to-host bandwidth issues.

Again, it all depends on the details of your simulation.

You might also consider just building the machine in a 4U rackmount case, like this:

http://www.newegg.com/Product/Product.aspx…N82E16811182566

If you get 4x GTX 295’s in one machine, it’s going to be really loud; if you’re just going to be running simulations, you might as well just put it in the next room (or the server room, if you have one). Also, like Atomiktoaster wrote, you might be close to maxing out the circuit that your wall outlet is on, which should be less of a problem if you can put it in a room that’s on it’s own circuit (or again, in a real server room with beefed-up electrical circuits).

I just built a system with the P6T7 motherboard in the ABS 595 Canyon case, using the Thermaltake 1200W power supply. The case has enough expansion slots to accomodate the motherboard, etc.

It is more expensive than the other case you are considering, but it is very quiet (I am using the V8 cooler instead of the stock Core i7 cooler) and dit oes the job. The only quibble I have is that they should ship better screws for the power supply tray extension but other than that I’m happy. This is the first computer I’ve built from scratch since the 1980s, and the build itself took one afternoon with no major points of confusion.

Hello Zenpharaohs,

I’m actually building a similar system with the P6T7 inside the ABS 695 case (worth the money). My question is, will the Thermaltake 1200W handle 3x Tesla’s along with a Quadro Card efficienlty? I’m starting off with 1 Tesla but eventually want to have 3 Tesla’s and my graphics card. Also which DDR3 RAM did you go with?

Thanks!