CUDA and openCL Benchmarks

astrocrab
astrocrab
Joined: 28 Jan 08
Posts: 208
Credit: 429202534
RAC: 0

ah, i see now. do you know

ah, i see now. do you know why is it was reverted?

Holmis
Joined: 4 Jan 05
Posts: 1118
Credit: 1055935564
RAC: 0

RE: ah, i see now. do you

Quote:
ah, i see now. do you know why is it was reverted?

Check this thread in "Problems and Bug Reports"

Jord
Joined: 26 Jan 05
Posts: 2952
Credit: 5779100
RAC: 0

RE: Sapphire HD7870 2GB,

Quote:

Sapphire HD7870 2GB, Windows 7 64bit Ultimate SP1, Catalyst 13.6-Beta2.
HD7870 on average over 9 tasks, 10147.07 seconds.

i5-2500K, BOINC allowed to use 3 cores for Seti v7, last core free for GPU. No overclocking. Only one task per core and GPU.

BOINC 7.1.18


All tasks done by HD7870 only (original 'broken one' and replacement), 18 tasks total time 187,807.05 seconds, average 10,433.73 seconds.

Now, I see the same artifacts on the present HD7870 as I did on the one I took back to the shop. It appears this is a problem with the 2D clocks of GPU and memory. Reading around other forums, all 7xxx cards have this problem, doesn't matter what brand you have. When the 2D memory goes into powersaving mode, it downclocks to ~150MHz and use very little electricity at that time, which make it problematic for the GPU to update 2D (text) screens, which in turn causes flickering of the screen and other artifacts.

Seeing how OpenCL only uses the OpenGL processors for calculations, but don't actually output anything 3D, the 2D clocks can still go into power-save mode when OpenCL runs. A way around this is to overclock the 2D clocks, bring the memory speed up to 400/500Mhz. But that's a bit difficult with the present MSI Afterburner, Asus GPU Tweak or even Sapphire TRIXX applications, as these only show the 3D clocks. I'll figure out a way. :)

Oh, by the way, temperatures have managed to find their way down now as well. Only 54C when running Einstein BRP5 in OpenCL (as opposed to 66C that the 'broken one' did), and 70C in Crysis3 (as opposed to 83C for the other one).
I also found a little redeemer card for Bioshock Infinite, Tombraider and FarCry3 Blood Dragon in the box the whole thing came in. Not too interested in TR, nor in FC3BD, but wondering about Bioshock. And that's the only game for which the game keys aren't ready yet. :-(

Jord
Joined: 26 Jan 05
Posts: 2952
Credit: 5779100
RAC: 0

LOL, and then the Einstein

LOL, and then the Einstein admins changed the playing field. So now the tasks are shorter.

BM wrote:

We decided to reduce the "bundle size" of BRP5 from 3 to 2, mainly to help people with extremely long run-times (which is pretty bad for a GPU App). This means that the run-time of new BRP5 tasks should be reduced to 2/3 of previous tasks. FpOps and Credit have been adjusted accordingly.

BM

So, my first 10 tasks at v1.34 with 2/3s the time of the previous run:
68,500.85 total time, thus 6,850.09 seconds average. Which is better than the run time the admins had their eye on, as (((187,807.05 seconds / 3) * 2) / 18) = 6,955.82 seconds. ;-)

Alinator
Alinator
Joined: 8 May 05
Posts: 927
Credit: 9352143
RAC: 0

LOL... Agreed. Just when

LOL... Agreed.

Just when you think you have all the answers, they change all the questions. ;-)

Or as Bones famously stated, "...I know engineers, they love to change things!". :-D

Sid
Sid
Joined: 17 Oct 10
Posts: 160
Credit: 921076823
RAC: 285954

Did anyone try ATI 7950 for

Did anyone try ATI 7950 for BRP tasks? It is cheaper that 7970 and theoretically can be overclocked to almost as 7970. Top computers in statistics have 7950/7970 in description so it is hard to say is it exactly 7970 or not.

Mumak
Joined: 26 Feb 13
Posts: 325
Credit: 3292754431
RAC: 1510793

I tried that on my 2x7950

I tried that on my 2x7950 (http://einsteinathome.org/host/7192129/tasks), but the tasks got stuck and didn't finish, so I had to abort them all. It's strange, since I have no issues with that machine and running MW@H (and WCG in the past) at full power there...
I'll give it one more try now.

-----

Mumak
Joined: 26 Feb 13
Posts: 325
Credit: 3292754431
RAC: 1510793

So this time it worked, you

So this time it worked, you can check the above link for results.

HD7950, 1000 MHz, 1 WU/GPU, GPU usage ~75%
When CPUs loaded with tasks: 5200-5700 secs (PA0006)
When no other CPU tasks running: 4250 secs (PA0006)

When comparing with other results note, that some of the WUs processed were of different computation size (PA0006: 450000 GFLOPs, PA007x: 680000 GFLOPs).

-----

Sid
Sid
Joined: 17 Oct 10
Posts: 160
Credit: 921076823
RAC: 285954

RE: So this time it worked,

Quote:

So this time it worked, you can check the above link for results.

HD7950, 1000 MHz, 1 WU/GPU, GPU usage ~75%
When CPUs loaded with tasks: 5200-5700 secs (PA0006)
When no other CPU tasks running: 4250 secs (PA0006)

When comparing with other results note, that some of the WUs processed were of different computation size (PA0006: 450000 GFLOPs, PA007x: 680000 GFLOPs).


Thank you for the information. So 7950 is not much slower that 7970 ?

Mumak
Joined: 26 Feb 13
Posts: 325
Credit: 3292754431
RAC: 1510793

One more result: PA0070 WU

One more result:
PA0070 WU (this is the larger size), CPUs partially loaded with tasks: 8050 secs

-----

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.