I just joined and have a few questions:

What is the performance difference between a Tesla M2050/70 and a GTX580 in crunching here at GPUGRID. Between the 2050 and the 2070 does it make a real difference for crunching.

I plan two install two boards in one desktop. I can have two GTX580 no problem, I know this will work. But can I have instead two M2050/70. And if I have two Tesla boards is also the standard video display function available (like in the GTX580) or I need an additional video card.

At the moment I started by running as a test an Asus Matrix GTX285 board with the CPU running on WCG tasks. It seems that all works fine between the two applications. I get only one GPUGRID WU running at a time and it runs for nearly 10 hours on the GTX285. Is this normal, good or bad performance?

Many thanks for your help

After looking at existing threads it seems that the Tesla cards at least the previous ones 10XX series were not performant against the gaming cards of the Fermi type 400 series. Does this hold true even for the new ones of the 20XX series ?

For the GPUGRID application, M20?0 cards perform about as well as a GTX470. Not that we've had one to test yet, but a GTX580 should be ~25% faster.

MJH

Ok thanks MJH. It is clear that Tesla is not the right way here. Nevertheless I am very very surprised. I understood that Tesla was the Nvidia strategy for supercomputing, massive number crunching approach. How come that the gaming boards are so much better? I understand that the design of the Fermi boards was made with a number crunching perspective but the Tesla should have been 50% more performing than the gaming boards not the opposite.

After the excellent behaviour of the test machine with both GPUGrid and WCG crunching split between CPU (i7 980X running at 4.0 Ghz) and GPU (Asus GTX285 no OC) I have added two more machines that were crunching with WCG. They are all with 6.10.58 64 bit boinc release. The machines are all Win 7 64 bit and have 980X CPUs. Same config but with an Asus GTX 280 and an Asus GTX 260. These were older boards I had anyway and were not really used. This makes three boards and three machines connected. If all goes well I may add one or two 580 boards. I refused to buy the 480 as I consider it an unfinished board. The 580 is the one that should have been released first.

Hypernova, you are very welcome to GPUGrid.

Your 200 series cards should do reasonably well. On Windows stick to the 19745 driver or earlier for this card series; it allows you to run the faster 6.12 app (for now). For Fermi cards use the latest driver (260.99 for 400 series, 262.99 for 500 series) for Win.

Linux x64 is the best operating system here, (x86 versions will not work), closely followed by Win XP x64 but an x86 Win client is not too far off (within 3%). Vista and Win7 are about 11% slower than XP, after optimization.

The GTX580 is the top NVidia gaming card and the best card for crunching here. Stay away from the Tesla cards if buying.

We can advise on performance if task links are included in posts. You can PM here if you dont want to show devices in public. It's usually a good idea to up the fan speed on your cards (I generally use EVGA Precision because its easy to use).

Good luck,
Kev

Thanks skgiven, you have impressive numbers here at GPUGrid. The reason I came here is that I feel GPU processing at WCG is too far distant, and to leave some good GPU boards asleep is a pity. Some of my devices were designed to handle GPU processing for WCG, but as CPU and GPU both can run together without trouble then let's go. The projects that run here at GPUGrid have also attracted my interest.

At the moment all the cards run the 260.99 Nvidia driver. What do you mean when you say that the 197.45 driver would allow to run the faster 6.12 app. Do you mean WU's with shorter runtime, or a better use of the capability of the 200 series card which result in faster crunching the same WU.
Otherwise I do not mind that a WU lasts 10 or more hours as long as it is run efficiently.

Regarding the Win version it is a pity that W7 does not do better than XP but for me it is a no brainer. All my machines are standard on W7 64 bit Ultimate and this will not be changed. Regarding Linux I did a try at WCG but run into many problems and it was a completely new world for me. It would have needed too much effort to become literate on the subject and have things up and running as I like, so I gave up.

The 6.12 app uses the capability of the 200 series cards better (it's faster), but the cards require the 19745, or slightly earlier driver, to run this faster app. All the WU's are basically the same but modified to run on specific apps, and allocated according to your driver. So you can presently run 6.12 or 6.13 if you use 19745 or a later driver respectively.

To get some idea of speed, my GTX260-216 on Linux at stock (1.35) running app v6.12:

TONI_MSM 14719sec = 4.09h, 24ms per step. (625000 steps)
KASHIF_HIVPR 21901sec = 6.08h, 22ms per step (1000000 steps).
GIANNI_DHFR500 21901sec = 6.22h, 11ms per step (2000000 steps) - this is the fastest task type.

- runtime depends on matching the app to your card, the task length (steps), and task type (project); some use faster algorithms.

From what I've seen, the main advantage of the Tesla boards is that they offer more graphics memory. However, the GPUGRID applications were designed so they have no real use for this extra graphics memory.

From what I've seen, the main advantage of the Tesla boards is that they offer more graphics memory. However, the GPUGRID applications were designed so they have no real use for this extra graphics memory.

I would think that Tesla would be significantly faster for MilkyWay though since the DP units aren't crippled and now that the MW client is working with Fermi. Still nowhere near the performance of a MUCH less expensive ATI gaming card.