CUDA and openCL Benchmarks

k6xt
k6xt
Joined: 11 Sep 05
Posts: 7
Credit: 28,650,981
RAC: 2,262

Many thanks. I don't know if

Many thanks. I don't know if I have a PCIE3 board, will check. Haven't built a new PC in over a year, something to do this winter.

What is "the right CPU"? I have Q6600's and Q9650.

Regards
Art

Sunny129
Sunny129
Joined: 5 Dec 05
Posts: 162
Credit: 160,342,159
RAC: 0

its not that your particular

its not that your particular CPUs don't support the PCIe 3.0 standard so much as it is the fact that none of the chipsets that support intel Core 2 Quad CPUs (like yours) were PCIe 3.0 capable. i believe the best thing available back when C2Q CPUs were still on the shelf was PCIe 2.0. in short, you won't find a socket LGA 775 mobo with PCIe 3.0. for that, you'll have to step up to a mobo with a newer chipset, and hence you'll need to upgrade the CPU too.

Jeroen
Jeroen
Joined: 25 Nov 05
Posts: 379
Credit: 738,423,020
RAC: 0

RE: Many thanks. I don't

Quote:

Many thanks. I don't know if I have a PCIE3 board, will check. Haven't built a new PC in over a year, something to do this winter.

What is "the right CPU"? I have Q6600's and Q9650.

Regards
Art

In case you plan to build a new system, the following desktop processors support PCI-E 3.0.

LGA1150 Haswell 4th generation i3, i5 and i7
LGA1155 Ivy Bridge 3rd generation i5 and i7 *
LGA2011 Ivy Bridge Extreme 3rd generation i7 ***
LGA2011 Sandy Bridge Extreme 2nd generation i7 **

* If the motherboard has PCI-E switches to split lanes across slots for example for SLI, the switches must be rated for 3.0. Newer Z77 boards generally have this.
** For NVIDIA cards, a NVIDIA GEN3 patch has to be run for PCI-E 3.0 support since the drivers default to PCI-E 2.0. AMD card drivers enable PCI-E 3.0 support by default.
*** The same as ** until NVIDIA updates their drivers to where the drivers will default to PCI-E 3.0 without the GEN3 patch.

The 2011 CPUs have 40-lanes and the 115x CPUs have 16 lanes. There are different variant of the Xeon processors including E3 and E5 processors 2nd generation and newer that support PCI-E 3.0. The 2nd generation LGA2011 Xeon processors still require the above NVIDIA GEN3 patch for 3.0 support via NVIDIA cards even though Intel has them listed as supporting 3.0.

Upcoming AMD FM2+ based processors are supposed to support PCI-E 3.0 with 20-lanes.

David Rapalyea
David Rapalyea
Joined: 3 Jan 13
Posts: 79
Credit: 63,886,821
RAC: 0

FINAL PRODUCTION REPORT

FINAL PRODUCTION REPORT (multicard GPU)

Production RAC: 80,224
Power Draw At Plug: 220-225 Watts
10,000 stones/day: 28 Watts (using 225 power draw)

Machine Name: UserPC
Processor: i3-3220 @3.3gh
MotherBoard: ASRock Z77WS
PCIE 3.0 Slots: 4
GPU: 3xAsus GTX 650 cordless
PSU: XXX 350 Gold

I am thinking of cloning this machine but I am afraid to unplug it and jinx the setup....

Arecibo 19 Oct 2012
Just Because The Space Alien Is Green
Does Not Mean You Should Go

k6xt
k6xt
Joined: 11 Sep 05
Posts: 7
Credit: 28,650,981
RAC: 2,262

Interesting. Each of the PC's

Interesting. Each of the PC's completing this WU use 560ti's which appear to have equal specs. Yet my WU partner completed the unit in about half the time.

My PC is this one:
http://einsteinathome.org/host/5419234.
The work unit:
http://einsteinathome.org/workunit/174493049

I'm running at count 0.5, perhaps that's part of it. But even by subtracting out the difference in CPU time the other GPU is a good bit faster.

Maybe I should try running just one GPU WU at a time? Looking at the result it didn't seem like my card was overtaxed.

Regards
Art

mountkidd
mountkidd
Joined: 14 Jun 12
Posts: 127
Credit: 4,112,867,754
RAC: 2,238,797

RE: Maybe I should try

Quote:
Maybe I should try running just one GPU WU at a time? Looking at the result it didn't seem like my card was overtaxed.


Hi Art,
There is no simple or easy answer, but your card is most likely undertaxed.

Looking at your wingmate's info, he like you is running a full complement of E@H apps. But we can't tell if he is overclocking cpu or gpu, we don't know how many simultaneous gpu tasks are being run and we don't know if he is running other projects.

Running cpu tasks impacts gpu performance. Running a mix of different gpu tasks also impacts gpu performance. The only way to assess the impact on gpu performance is through controlled testing.

Turn off all cpu tasks. Select one type of gpu task and config for 1wu. Run enough wu's to get a meaningful metric. Config for 2wu and repeat. Config for 3wu and repeat. Config for 4wu and repeat. Monitor cpu & gpu temps and be sure to keep temps under control. Once you get through gpu testing and have decided on a config, you can add in a cpu task at a time to assess the impact on the gpu.

You may find that simpler is better.

Gord

Sid
Sid
Joined: 17 Oct 10
Posts: 146
Credit: 492,655,600
RAC: 302,638

RE: Looking at your

Quote:

Looking at your wingmate's info, he like you is running a full complement of E@H apps. But we can't tell if he is overclocking cpu or gpu, we don't know how many simultaneous gpu tasks are being run and we don't know if he is running other projects.

Gord

Actually it is possible to say that your wingmate runs only one WU. According to the log:
[10:32:50][8184][INFO ] CUDA global memory status (GPU setup complete):
------> Used in total: 414 MB (611 MB free / 1025 MB total) -> Used by this application (assuming a single GPU task): 210 MB

This is first gpu and Windows need some video memory too.

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,312
Credit: 46,547,204,406
RAC: 34,276,412

RE: Interesting. Each of

Quote:
Interesting. Each of the PC's completing this WU use 560ti's which appear to have equal specs. Yet my WU partner completed the unit in about half the time.


As has already been mentioned, your wingman is running tasks singly, so your 560Ti is actually slightly outperforming the other one even although your FX-4100 CPU is quite underpowered compared to the i7-2600 in the other host. By comparing the CPU component in both cases you can get an impression of how much better again your runtime would be if your 1400 sec value could be lowered to something closer to the 400 sec value of the i7. You are gaining a considerable advantage as it is, just by running 2x. You would probably get a further (but smaller) advantage by running 3x (count = 0.33).

I run a couple of hosts with 550Tis. They perform a bit better on 3x. They also run CPU tasks on all cores. There was no significant gain by freeing a CPU core. As has been mentioned, because of the variables involved, the only way to find out for sure is to experiment for yourself.

EDIT: One last point I should mention. If you attempt to max out your card, make very certain your cooling is up to the task.

Cheers,
Gary.

Sunny129
Sunny129
Joined: 5 Dec 05
Posts: 162
Credit: 160,342,159
RAC: 0

RE: You are gaining a

Quote:
You are gaining a considerable advantage as it is, just by running 2x. You would probably get a further (but smaller) advantage by running 3x (count = 0.33).


although i now run a pair of GTX 580s, i used to have a pair of GTX 560 Ti's...so i can confirm that 3 simultaneous E@H tasks are more efficient than 2 simultaneous E@H tasks, at least on a Win7 x64 platform.

Shafa
Shafa
Joined: 31 May 05
Posts: 53
Credit: 627,005,014
RAC: 0

RE: although i now run a

Quote:

although i now run a pair of GTX 580s, i used to have a pair of GTX 560 Ti's...so i can confirm that 3 simultaneous E@H tasks are more efficient than 2 simultaneous E@H tasks, at least on a Win7 x64 platform.

I have been testing 560ti on WinXP 32bit, Xubuntu 12.04 32bit and Debian Wheezy 64bit since last week (ok, not with the latest drivers) and there is no advantage to have 3 parallel tasks instead of 2 ones on this card.

Edit:
Performance depends also on other factors - cpu&ram (kind, speed), PCIe speed (16x vs 8x), etc.
So there may be different experience even with the same gpu.

Comment viewing options

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