Welcome to MilkyWay@home

Speed of ATI-App


Advanced search

Message boards : Application Code Discussion : Speed of ATI-App
Message board moderation

To post messages, you must log in.

AuthorMessage
perplexityCG

Send message
Joined: 11 Sep 09
Posts: 1
Credit: 0
RAC: 0
12 year member badge
Message 34439 - Posted: 10 Dec 2009, 16:00:06 UTC

Hi everyone

Recently the calculating time of my WUs with my GPU was quadrupled, I assume that's going with the quadrupling of the Credit per WU.
However, at all the speed seems anyhow tremendously low.
My CPU is a Core i7 @ 3.8 GHz, that crunches now as fast as my HD 3870, in particular that means 8 WUs with HT @ 40 minutes per WU. According to Sandra my CPU achieves around 80 GFLOPS while my GPU should have 116 DP-GFLOPS (OC by 10%). So maybe the CAL-App may be not optimized as well as my CPU-App, no idea.

BUT: You see, my acc has several computers. One of it (InneB4 with Linux and a Phenom II X4 955) recently acquired a GeForce 9600GSO. According to CUDA-Z, he has 330 SP-GFLOPS, so he compiled the App with SP-Support. Now he crunches a WU, which takes usually 8.5 minutes with my HD 3870, in 0.5 minutes. How is that possible? Where does this difference in the performance come from?
In the ATI-App preview I saw the screen of Cluster Physik with units calculated within 6-9 seconds. How was that possible? A HD 4870 has approximately a 120% increased performance compared to a HD 3870.

Anyone an idea where this comes from? I'm quite curious. Tank you.
ID: 34439 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileGary Roberts

Send message
Joined: 1 Mar 09
Posts: 56
Credit: 1,984,937,499
RAC: 0
1 billion credit badge12 year member badge
Message 34461 - Posted: 11 Dec 2009, 13:11:26 UTC - in response to Message 34439.  

Recently the calculating time of my WUs with my GPU was quadrupled, I assume that's going with the quadrupling of the Credit per WU.
However, at all the speed seems anyhow tremendously low.

There is no change in speed. The work content of a task has been quadrupled - check out the front page news item.

My CPU is a Core i7 @ 3.8 GHz, that crunches now as fast as my HD 3870, in particular that means 8 WUs with HT @ 40 minutes per WU. According to Sandra my CPU achieves around 80 GFLOPS while my GPU should have 116 DP-GFLOPS (OC by 10%). So maybe the CAL-App may be not optimized as well as my CPU-App, no idea.

I don't think it's wise to try to compare a CPU with a GPU in this way. In any case, the ATI GPU app is very efficient.

BUT: You see, my acc has several computers. One of it (InneB4 with Linux and a Phenom II X4 955) recently acquired a GeForce 9600GSO. According to CUDA-Z, he has 330 SP-GFLOPS, so he compiled the App with SP-Support.

If you are saying that you have compiled a single precision app so that you can run on a single precision GPU, you are really breaking the project requirements very badly. The project requires double precision - please read this thread and in particular messages like this one and then please stop returning useless inaccurate results. I'm surprised that the validator is allowing the results to be marked valid and awarding you credit.

Now he crunches a WU, which takes usually 8.5 minutes with my HD 3870, in 0.5 minutes. How is that possible? Where does this difference in the performance come from?

Well of course it will be a hell of a lot faster if you are only using single precision. Your 3870 is obviously using a correct app ie. double precision!

In the ATI-App preview I saw the screen of Cluster Physik with units calculated within 6-9 seconds. How was that possible? A HD 4870 has approximately a 120% increased performance compared to a HD 3870.

Anyone an idea where this comes from? I'm quite curious. Tank you.

Did you also notice the date on that screen shot? It was almost a year ago when the size of a task was very much smaller than what it is today. There is absolutely no puzzle about this.

Cheers,
Gary.
ID: 34461 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote

Message boards : Application Code Discussion : Speed of ATI-App

©2021 Astroinformatics Group