CUDA performance

[asadf4444444444]

From another thread: 9800GT GPU only??? = 14000

http://www.overclockers.com/forums/showpost.php?p=6401944&postcount=5

That chart is for Total credits not RAC.

Kryten451, is this you?

 

[Kryten451]

That chart is for Total credits not RAC.

Kryten451, is this you?

Yes, that's me and my stats with 9800 GT so far

 

[QuietIce]

That chart is for Total credits not RAC.

Well that should have been obvious!

Yes, that's me and my stats with 9800 GT so far

Care to push the "Show host average" button (on the left) and post it ...???

 

[Kryten451]

Ahh, Show user average, So this would be my RAC ..

 

[QuietIce]

"Show user average" would be your RAC for all machines you run - since you're the user.
"Show host average" would give a graph for that machine only.

Of course, if you're only running one machine (?) then it would be that same graph ...

 

[Kryten451]

Yes I just have the one machine.. buy my Numbers should roughly triple when I deploy the new system below during the week having 328 stream processors total...

Stream processors 216(260GTX) + 112(9800GT)

 

[QuietIce]

When your 9800GT finally tops out would you please post it's RAC so I can include it with the CUDA RAC database ...?

 

[Kryten451]

When your 9800GT finally tops out would you please post it's RAC so I can include it with the CUDA RAC database ...?

Unfortunately I had to dismantle that old host during the Berkley Air Conditioning down period. I will build it up again with an identical 9800GT.. and post the host totals when RAC tops-out.

 

[donuts]

As soon as everything stabilizes, I'll have a couple to add to the database. GTX280 and a 9600gso.

 

[donuts]

Well, with the recent outages, I couldn't get real solid numbers and the cards are being pulled out of service.
#'s may be on the low side but fairly close.
gtx280-8500 rac
9600 GSO-3500 rac (96 sp)

The 280 was in my daily rig, 9600 in a cruncher.

 

[QuietIce]

Thanks! for the input!


I added the following lines to the list at the top. If something is incorrect or there's anything else to add (i.e., dedicated CPU cores for the GPU, changes in speed, etc.) please let me know.

9600GSO (G92) ~3500 96:48:12 @ 550 (dedicated cruncher)
GTX280 (GT200) ~8500 240:80:32 @ 602

 

[donuts]

Just the GTX has 240 shaders. I believe the rest is good.

 

[QuietIce]

Oops! Fingers wandered on the 10-key! *redface* 2-4, 3-5; same sequence just 0.5" left ... LOL!


Corrected - Thanks!

 

[Voodoo Rufus]

Trying the trick of leaving one CPU core free to talk with the CUDA cards. I'll give a performance estimate of my eVGA GTS 250 Superclocked cards in the near future.

 

[Duner]

Just a teaser. This GTX260 192sp was doing 6000 RAC when I was crunching Rosetta on the CPU, now it's over 9000 RAC with the CPU idle. Still, climbing a bit, I'll post back when it's stabilized. Definitely seems like A spare CPU benefits the GPU. The question still remains how many spare CPU's do you need for a multi-GPU setup.

http://setiathome.berkeley.edu/show_host_detail.php?hostid=5130705

 

[Voodoo Rufus]

Running two GTS 250s on my QX9650 with one core freed gives between 80-90% total CPU utilization, or about 25% usage on the free core for Cuda.

 

[QuietIce]

Is there any way for you to estimate the RAC on those two cards?

What speed are you running the QX9650? I have no clue what a QX9650 runs in RAC by itself but maybe someone does so we can make an educated guess at the video cards?

 

[Voodoo Rufus]

The QX9650 at 3.6GHz on all four cores will average 7k RAC optimized.
A GTS 250 puts out about 42% more FLOPs than a 9800GT. Your list shows those at about 3k rac, so mine should be around 4200 each with the factory overclock. With one core disabled on the QX it should level around 5250 by itself. Total RAC for the computer with the cards might end up around 13650.

 

[QuietIce]

We'll wait a couple of weeks to see where it ends up then - no hurry. Didn't realize you had just switched over.


BTW
The DAC for that rig is already at 13650 - when did you switch over to three-core crunching ...?

 

[FloridaBear]

I've seen it recommended to leave one CPU core free to "manage" the CUDA process. Is this primarily an issue with the priority of that process? Wouldn't it improve overall throughput to make that CUDA process "below normal" instead of "low" while still using all cores? I assume BOINC does not dedicate a core to this CUDA process. But, if it just gave it CPU time when needed, it should run nearly as fast as if it had a dedicated core.

I could likely write a service that would simply monitor for this process and change the priority when it finds that process out there.

Thoughts?