I know that this has been touched on differently in a couple of other threads, but I couldn't find a direct comparison like the following...

I recently got the opportunity to stress test a machine with a Quadro FX 3700 (basically an 8800GT/9800GT G92 core with the extra Quadro professional elements added). Since it has 112 shaders, I figured that it would out perform my 9600 GSO (also G92) with only 96 shaders. The other major differences between the cards are:

1. usage (the 9600 is on a regularly used machine, the Quadro is on a machine running for a few days to show it is stable--no other use)

2. The Quadro has 512mb, the 9600 GSO has 384mb

and

3. the 9600 is factory OC (reports here as 1674000) while the Quadro is stock and slower (1242000)

The 9600 GSO tends to report workunit elapsed time around 72,000 sec (19 - 20 hours) with low 80's for time/step. The FX 3700 is actually a bit slower with elapsed times in the 80,000 sec range (22 - 23 hours) and time/step in the low to mid 90's.

I was surprised by this. I figured that the faster clock on the 9600 would help it compensate for the disadvantage in shaders, but I didn't think that it would actually be the faster card in this situation.

I am curious if 8800GT and/or 9800GT cards with different clocks are consistent with this finding, or if the Quadro is itself structured differently enough to cause this?

I know that this has been touched on differently in a couple of other threads, but I couldn't find a direct comparison like the following...

I recently got the opportunity to stress test a machine with a Quadro FX 3700 (basically an 8800GT/9800GT G92 core with the extra Quadro professional elements added). Since it has 112 shaders, I figured that it would out perform my 9600 GSO (also G92) with only 96 shaders. The other major differences between the cards are:

1. usage (the 9600 is on a regularly used machine, the Quadro is on a machine running for a few days to show it is stable--no other use)

2. The Quadro has 512mb, the 9600 GSO has 384mb

and

3. the 9600 is factory OC (reports here as 1674000) while the Quadro is stock and slower (1242000)

The 9600 GSO tends to report workunit elapsed time around 72,000 sec (19 - 20 hours) with low 80's for time/step. The FX 3700 is actually a bit slower with elapsed times in the 80,000 sec range (22 - 23 hours) and time/step in the low to mid 90's.

I was surprised by this. I figured that the faster clock on the 9600 would help it compensate for the disadvantage in shaders, but I didn't think that it would actually be the faster card in this situation.

I am curious if 8800GT and/or 9800GT cards with different clocks are consistent with this finding, or if the Quadro is itself structured differently enough to cause this?

Remember that is elapsed time, if there is anything else running that might slow down the cpu's ability to communicate with the gpu, this effects elapsed time, meaning part of that time the gpu could be idle while it waits on the cpu which is busy doing something else.

I have two nearly identical systems, hardware and speeds, with only a few software differences. The one I use more heavily shows about 2,000 sec more elapsed time than the other. They even show this when I do not use them, like letting them run over the weekend.

But that is just it...the 9600 GSO is in an old Pentium D 830 in my office which is in regular use including BOINC CPU tasks 24/7, but the Quadro FX 3700, which was slower on the GPU task, is in a new Q9450 quad core which was only running BOINC CPU tasks along with the GPU unit (Indeed, I was able to put BOINC on it temporarily to max out the CPU and GPU to test what I and the IT person setting it up thought was a questionable 375 watt power supply)?

Has anyone else noticed that credit for longer processed WU on slow GPU's is less than credit issued for fast GPU's even thought the WU was issued to only one computer. I believe that credit issued should be the same for every WU irregardless of what GPU it was processed on. If it was completed before the 4 day deadline then full credit should be issued!!!

Pat

The credit are fixed for each WU regardless of the GPU used. It just take longer to finish it.

gdf

Please have a look at my completed WU's. The 6000-8000's are on the 280's and the 30000-40000's are on my 8800's. You can plainly see that the same amount of credit is not issued and most of the WU's were only sent to my computer. Don't now what to say?

Pat

That's the cpu time, not the elapsed time.

gdf

Yes, I know but if you look at the claimed credit and granted credit for both my rigs on the last 20 WU the values change. I was under the assumption that credit per WU is 3,232.06. The first WU that wasn't, was WU 117626 sent out 10 Dec and rtnd 11 Dec. Since then my claimed and granted credit has been fluctuating on both my rigs so much that I get 2,435.94 and 1,887.97 as common credits. It is something that I just noticed today but seems to have started happening about the same time these DCF and no WU issues started. Any possible correlation?

Pat

Those should be different version workunits. The higher credit for version 6.52 and others; the lower credit for 6.55--it is a different kind of workunit.

Sorry, don't want to burst your bubble but; task 121076 got 3232.06 credit, task 126125 got 2435.94 credit, and task 126252 got 1887.96 credit and they all ran under application version ACEMD 6.55. Any ideas GDF?

Pat

Well, you are correct. Hadn't seen a longer 6.55 until that one. I thought that it matched up with versions, but on that I am wrong. I would still maintain that they are different types of units doing slightly different sets of calculations (or maybe debugging code). Essentially, all the 3200 credit workunits have longer elapsed times with shorter time step than either the 2400 credit units or the 1800 credit units. That would seem to suggest additional calculations per step in the lower credit/shorter workunits.

See reply 4567 under Nvidia GPU: What's going on with the credits?

We need to decide where this thread should carry on. I believe we should keep it here.

GDF, balls in you court!

Cheers

Pat