Four GTX-580 on a single consumer board

Yes, I agree. It was a bold extrapolation.

Summing up, it is probably better for me to stick for the moment to the two GTX-580 I have. No benchmarks for molecular dynamics have been carried out so far for the 600 series (at least, no one came out from the forums) and those on game/OpenCL are not uniform. Also, probably a motherboard with x16 PCIe 3.0, not my x16 PCIe 2.0 will be needed to get most from higher cards than GTX-580.

Finally, I was impressed from the performance and (comparatively low) price of the fastest Radeon. There is a molecular dynamics MD European (opensource) code (European) that runs on ATI-Radeon, currently with OpenMM, which is said hard for programmers. Although the code is available on my Debian amd64, in the past, at the time of CPU only, I found difficult (from the chemical point of view) to work with such MD code. They are examining the possibility to shift to OpenCL, as I heard from them "The ATI hardware can process a lot more operations in parallel but those operations are slower. NVidia processes fewer operations in parallel but those operations are faster. Differences in hardware architechure requires changes to how certain problems can be programmed and final throughput is not the same thing as the peak GFLOPS, depending both on the problem and on the program." In contrast with US MD code, where shift from GPU to CPU is very frequent, the European code stays longer at GPU.

thanks for all kind advice
chiendarret

Originally Posted by chiendarret Also, probably a motherboard with x16 PCIe 3.0, not my x16 PCIe 2.0 will be needed to get most from higher cards than GTX-580.

The GTX580 doesn't support PCIe 3.0.

I agree. Although it is what people of the MD (molecular mechanics) in the US say, it was a bold extrapolation from my side.

All together, moving to GTX-680 or 690 with my GA-890FXA-UD5 motherboard is risky. Perhaps, the motherboard cannot exploit all the potential of the cards, while no benchmark has appeared with for the 600 series in MD. I plan to wait until motherboards with two x16 3.0 are available in the consumer market, and the 600 series is tried with MD, unless I can move to ATI-Radeon.

What emerged here is that with ATI-Radeon I could save much money, while perhaps running faster. Actually the European MD code (available as a Debian amd64 package, which is just my OS), runs on ATI-Radeon cards, programmed on OpenMM. As the latter is said to be hard in programming, the developers plan to move to OpenCL. That European code is of excellent quality, albeit limited for certain type of molecules. If I can find the way to manage the latter issue, a move to the European MD code is the best perspective.

Thanks
chiendarret

Originally Posted by chiendarret As a final question, sticking as I must to two x16 lanes, 2.0

Why does it have to be PCIe 2.0 x16?

If is load balancing, you might be able to set the BIOS to x8/x8/x8 (and from the gigabyte website, I'd say that the the default if you install 3x PCIe x16 cards).

Originally Posted by chiendarret I forgot previously to answer why CPU/GPU = 2. In molecular dynamics the largest part of the computation (the non-bonded) is carried out by the GPU, and this is why the GPU proved so useful, provided that much job is requested. Still, energy calculations are left to the CPU because it proved difficult to implement this part to the GPU. very recently, developers are beginning to overcome this obstacle.

Different CPUs and GPUs have very different capabilities.

You could have setup like this- Athlon 64 X2 3600+ (dual core) and GTX 580. Or Core i3 2130 and 9400GT. Both systems are 2 x CPU cores, single GPU...both would behave _very_ differently (the athlon has far less CPU power, and far more GPU power).

2 x CPU cores pre GPU seems to be a very basic 'rule of thumb' and like most rules of thumb its not that hard to skew.

Without actually seeing some benchmarks, thats as far as I will go.

Originally Posted by chiendarret Summing up, it is probably better for me to stick for the moment to the two GTX-580 I have. No benchmarks for molecular dynamics have been carried out so far for the 600 series (at least, no one came out from the forums) and those on game/OpenCL are not uniform. Also, probably a motherboard with x16 PCIe 3.0, not my x16 PCIe 2.0 will be needed to get most from higher cards than GTX-580.

GTX 680 looks better on paper. In reality, for compue use the GTX 680 is not that much better, and in some cases weaker than the GTX 580, particularly for FP64. GTX 580 memrmoy bandwith is actually slightly higher than GTX 680. GTX 680 is much more optimised for gaming over compute than previous recent 'top of the line' nVidia GPUs. nVidia may actually be planning on splitting its 'gaming' GPUs from its quadro/'compute' GPUs.

There are boards around with (unoffical) PCIe 3.0 support. For gaming, it doesnt make much difference in general, and at high resolutions there is very little difference (and sometimes PCIe 2.0 x16 can be faster).

http://www.techpowerup.com/reviews/I...Scaling/1.html

How much difference there would be for your use...no idea.

I admit that it was a bold assumption from my side. That said, the panorama has become rather clear. The MD software I use splits - as I wrote - the computation between GPU and CPU, so that there is communication at every step. Thus, at PCIe 2.0 the risk exists that gtx680 (bandwidth 256-bit) comes out slower than gtx580 (bandwidth 384-bit). To go ahead, PCIe 3.0 (double transfer rate) is a must with gtx680 in my case. If so, with very large computation, 680 (more cores) could be faster than 580 (less cores).

Should ab initio computations become available on GPUs, 680 is inferior to 580. The latter has better capability for double precision.

On the other hand, all motherboards that I see with PCIe 3.0 have Intel sockets. This means for me paying a lot more for hyperthread that will not be used. Perhaps a former AMD Thuban 6-core could do even better that Phenom II.

In conclusion, could you suggest a not expensive motherboard with two x16 at PCIe 3.0, possibly with Phenom II, which I already have?

thanks
chiendarret

The current top of the line chipsets from AMD's 900 series don't support PCIe 3.0, so I doubt that you will find a board for AM3(+) with PCIe 3.0.