CUDA and openCL Benchmarks

Alex
Alex
Joined: 1 Mar 05
Posts: 451
Credit: 507044931
RAC: 101212

RE: I would compare my

Quote:

I would compare my overclocked 550Ti to my overclocked 750Ti for you, but both have been gone from Einstein for long enough that there isn't much evidence left in the statistics.

I've re-enabled Einstein on my GTX750ti, GPU utilization factor 1.
If you're interested you can follow the progress here: http://einsteinathome.org/host/6801076/tasks
Additional power consumption while running a CasA 1.08 app can be seen here:
https://dl.dropboxusercontent.com/u/50246791/screenshot.PNG
TDP is less than 70Watts.

Phil
Phil
Joined: 8 Jun 14
Posts: 579
Credit: 228493502
RAC: 0

FYI Computer ID

FYI

Computer ID 11394499

i3 530 2.93 Ghz Nvidia 750 (not ti)

Typical run time for a Perseus Arm Survey wu is 18k sec with 4.3k cpu time.
This is running 2 concurrent wu's, currently set to run Perseus Arm Survey and Arecibo GPU wu's.
This is with stock settings for entire computer and air cooling.

These times should improve with a better cpu and motherboard. You can go to the home link at the bottom of this page, scroll down and select statistics, then select top hosts. You can then scroll down and peruse hosts with various GPUs and you will find some 750s and 750ti's. Look at the supporting cpu's and memory, etc.

I really like the 750s and ti's. Good throughput at lower power.

Hope this helps.

Phil

Phil
Phil
Joined: 8 Jun 14
Posts: 579
Credit: 228493502
RAC: 0

The following is a comparison

The following is a comparison between the same exact cards but with computers having different specs. It is older gear but should give an idea of how differing computer specs can affect GPU throughput.

Both machines run Linux Mint 17, and Nvidia GT 630 GPUs.

11454899 host
Intel E5400 2.7Ghz 2Gig memory RAC 8839

11468802 host
Intel E2180 2.0Ghz 1Gig memory RAC 6032

As you can see, the supporting structure such as CPU and memory, motherboard, etc. have great influence on how much a GPU can process. These are the only 2 machines I have with the same GPU card in them, hence the only 2 I can do a straight up comparison for you.

That being said, I have an i3 machine with a 750 (non ti) and another E5400 with a 760 installed. The 2 machines are running neck and neck for RAC. You would think the 760 would be running away from the 750, but the 750 has a better computer to support it. Both machines are running approx. 30k RAC.

If I get some energy today I might swap the cards and let them run to see where the RAC ends up and I'll post the results.

Phil

Phil
Phil
Joined: 8 Jun 14
Posts: 579
Credit: 228493502
RAC: 0

RE: That being said, I have

Quote:

That being said, I have an i3 machine with a 750 (non ti) and another E5400 with a 760 installed. The 2 machines are running neck and neck for RAC. You would think the 760 would be running away from the 750, but the 750 has a better computer to support it. Both machines are running approx. 30k RAC.

If I get some energy today I might swap the cards and let them run to see where the RAC ends up and I'll post the results.

Ok, quoting myself there, so here's the new numbers after the card swap.

i3 machine now has the Nvidia 760:
Perseus Arm Survey time has dropped to 12k from 18k.
Arecibo times have dropped to 3k from 7k.

E5400 machine now has the Nvidia 750 (non ti):
Perseus Arm Survey time has increased from 18k to 24k.
Arecibo times have increased from 4k to 6k.

RACs on both machines will take a few weeks to adjust.

Phil

mikey
mikey
Joined: 22 Jan 05
Posts: 12680
Credit: 1839083536
RAC: 3919

RE: RE: That being said,

Quote:
Quote:

That being said, I have an i3 machine with a 750 (non ti) and another E5400 with a 760 installed. The 2 machines are running neck and neck for RAC. You would think the 760 would be running away from the 750, but the 750 has a better computer to support it. Both machines are running approx. 30k RAC.

If I get some energy today I might swap the cards and let them run to see where the RAC ends up and I'll post the results.

Ok, quoting myself there, so here's the new numbers after the card swap.

i3 machine now has the Nvidia 760:
Perseus Arm Survey time has dropped to 12k from 18k.
Arecibo times have dropped to 3k from 7k.

E5400 machine now has the Nvidia 750 (non ti):
Perseus Arm Survey time has increased from 18k to 24k.
Arecibo times have increased from 4k to 6k.

RACs on both machines will take a few weeks to adjust.

Phil

So the i3, quad core, is much faster then the dual core machine! That would seem to indicate alot of cpu processing going on for the Perseus Arm Survey units.

Mumak
Joined: 26 Feb 13
Posts: 325
Credit: 3518901814
RAC: 1617809

RE: That would seem to

Quote:
That would seem to indicate alot of cpu processing going on for the Perseus Arm Survey units.

Indeed, even though the CPU utilization doesn't show relatively high numbers, it seems both BRP GPU (BRP4G and BRP5) tasks seem to involve a lot of CPU cycles.

I for example have been running on an AMD FX-8350 with R9 280X. I was wondering why those tasks take so much time. Then I found the reason - the power scheme in Windows was set to balanced and since it's summer here, I'm running low load only (no CPU tasks, etc). So BRP tasks were the only ones running on that machine and I noticed the CPU was running at the lowest (idle) clock. This has caused that the CPU couldn't feed the fast GPU enough. Solution was simple - either run an additional CPU intensive task, or switch the power scheme to Max.

To provide more comparison on how CPU speed is affecting the processing of BRP, here are some times for BRP4G (single task):
Radeon R9 280X @ 1050 MHz + AMD FX-8350 ~ 1550 s
Radeon HD 7950 @ 950 MHz + Core i7-3820 ~ 1500 s
So even if the GPU is faster (280X), the slower CPU results in a slowdown.

-----

mikey
mikey
Joined: 22 Jan 05
Posts: 12680
Credit: 1839083536
RAC: 3919

RE: RE: That would seem

Quote:
Quote:
That would seem to indicate alot of cpu processing going on for the Perseus Arm Survey units.

Indeed, even though the CPU utilization doesn't show relatively high numbers, it seems both BRP GPU (BRP4G and BRP5) tasks seem to involve a lot of CPU cycles.

I for example have been running on an AMD FX-8350 with R9 280X. I was wondering why those tasks take so much time. Then I found the reason - the power scheme in Windows was set to balanced and since it's summer here, I'm running low load only (no CPU tasks, etc). So BRP tasks were the only ones running on that machine and I noticed the CPU was running at the lowest (idle) clock. This has caused that the CPU couldn't feed the fast GPU enough. Solution was simple - either run an additional CPU intensive task, or switch the power scheme to Max.

To provide more comparison on how CPU speed is affecting the processing of BRP, here are some times for BRP4G (single task):
Radeon R9 280X @ 1050 MHz + AMD FX-8350 ~ 1550 s
Radeon HD 7950 @ 950 MHz + Core i7-3820 ~ 1500 s
So even if the GPU is faster (280X), the slower CPU results in a slowdown.

That's one of my own 'tweaks' I do on my own machines as soon as I put Win7 on them, but I save it for the friends pc's I work on to speed it up later on when they start complaining that it is too slow. That and I swap them over to an SSD drive for the C: drive, setting up a backup program and drive in the process.

Sid
Sid
Joined: 17 Oct 10
Posts: 164
Credit: 969895863
RAC: 417025

Does anybody have any

Does anybody have any experience with NVidia GTX 295? This two heads monster uses 896-bit (448-bit x 2) access to video memory. However type of memory is DDR3.
I can get this card for almost free so I'd like to get some information if it makes sense to use this card.

Alex
Alex
Joined: 1 Mar 05
Posts: 451
Credit: 507044931
RAC: 101212

RE: Does anybody have any

Quote:
Does anybody have any experience with NVidia GTX 295? This two heads monster uses 896-bit (448-bit x 2) access to video memory. However type of memory is DDR3.
I can get this card for almost free so I'd like to get some information if it makes sense to use this card.

I had a GTX295, gave it away two years ago. Works fine with cuda32 wu's, but you will most likely have troubles with cuda 5 and cuda 6 apps.
And it produces a lot of hot air.
Another thing to be aware: there were two versions, one with less than 200 cuda cores and one with > 200 cuda cores. The first version had a firmware (or hardware?) bug which was never repaired, the later one woked fine. If you need more info, look in the gpugrid forum 5 years or so backwards. Keyword 'crippled'.

MAGIC Quantum Mechanic
MAGIC Quantum M...
Joined: 18 Jan 05
Posts: 1886
Credit: 1406474600
RAC: 1169649

Free is always nice but that

Free is always nice but that card won't help you here.

I am looking around right now for another GeForce card for host #7 and I see a cheap GeForce GTX 750 with 512 CUDA Cores for only $110

Maybe the 660Ti price will start dropping and I am checking the 550Ti's and 650Ti's for under $150

Comment viewing options

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