Two projects on one GPU?

Peter Hucker
Peter Hucker
Joined: 12 Aug 06
Posts: 244
Credit: 218,670,728
RAC: 682,306
Topic 219883

Can I force Einstein and Milkyway both to run on the same GPU at once? The reason I'm considering this is Milkyway is using the double precision part of the GPU, and Einstein uses the single precision part. Are these parts of the chip independant? Can I run both and get twice the work done?

Keith Myers
Keith Myers
Joined: 11 Feb 11
Posts: 747
Credit: 638,529,637
RAC: 1,234,496

Think it is doable.  I have

Think it is doable.  I have had both Seti and MW set up for 0.5 count at times and have had both the Seti task running alongside the MW task on a gpu.  Normally with 3 gpus running I would see either double Seti or double MW tasks on a card, but sometimes I would get a MW task to jump in with the running Seti task at times on a card.

 

Peter Hucker
Peter Hucker
Joined: 12 Aug 06
Posts: 244
Credit: 218,670,728
RAC: 682,306

I have given up on this.  In

I have given up on this.  In another forum Richard Haselgrove told me how to make it run 1 of each on a GPU, but it produced precisely zero increase in work throughput.  I guess the way those projects are coded, or the way my graphics card works, it can't do both at once.  As for more than one of the same project on a card, that doesn't achieve much either.  Milkyway is already maxing out the card, and Einstein is only at 67%, but adding more tasks only increases the GPU load slightly, and the CPU load astronomically, while the rate it completes tasks remains the same.  Something is probably wasting time swapping registers on the chips?

Keith Myers
Keith Myers
Joined: 11 Feb 11
Posts: 747
Credit: 638,529,637
RAC: 1,234,496

Yes, it always depends on the

Yes, it always depends on the science apps and a lot of times on the type of task.  Back in the day, you could run 3,4 or 5 Seti CUDA tasks on a card at the same time.  No more.  One task at a time is the most productive with the shortest run_times with the current apps.  Seems the MW volunteers can get away with multiples on AMD cards using the previous generation of cards with ease.

The best way I know to get more production increase is to add another or more gpus and run them 24/7.

 

Peter Hucker
Peter Hucker
Joined: 12 Aug 06
Posts: 244
Credit: 218,670,728
RAC: 682,306

I'm on a Radeon 7970 for MW. 

I'm on a Radeon 7970 for MW.  This is maxed out with a single MW task.  7970 is about the best for Boinc stuff (performance per price) as I can get them for £60 second hand.  The R9 280X is a bit better, I'll be getting those in the future.  It irritates me that most newer cards are mainly aimed at doing single precision floating point (I guess that's all that's needed for games, which is their main market).  MW needs a good double precision speed.

mikey
mikey
Joined: 22 Jan 05
Posts: 5,910
Credit: 544,121,970
RAC: 87,075

Peter Hucker wrote:I'm on a

Peter Hucker wrote:
I'm on a Radeon 7970 for MW.  This is maxed out with a single MW task.  7970 is about the best for Boinc stuff (performance per price) as I can get them for £60 second hand.  The R9 280X is a bit better, I'll be getting those in the future.  It irritates me that most newer cards are mainly aimed at doing single precision floating point (I guess that's all that's needed for games, which is their main market).  MW needs a good double precision speed.

This chart should help if you want the best dual precision gpu, the last column is the single precision first, then the dual precision percentage that card does. For instance your 7970 is FP64 = 1/4 FP32 meaning the dual precision rate is only 1/4 as fast as the single precision rate. While a 1080Ti is only at FP64 = 1/32 FP32.

http://www.geeks3d.com/20140305/amd-radeon-and-nvidia-geforce-fp32-fp64-gflops-table-computing/

Peter Hucker
Peter Hucker
Joined: 12 Aug 06
Posts: 244
Credit: 218,670,728
RAC: 682,306

Yes, I took one of those

Yes, I took one of those charts a while ago, stuck it into Excel, and added the cost of buying each card, then got it to calculate which gave the most bang for the buck.  Every so often I look at new cards to add to it, and update the prices.

This website gives the DP and SP of any card: https://www.techpowerup.com/gpu-specs/radeon-hd-7970.c296

mmonnin
mmonnin
Joined: 29 May 16
Posts: 265
Credit: 924,144,806
RAC: 275,056

I run MW on my RVII with E@H

I run MW on my RVII with E@H as my backup. During the periods where MW won't send tasks I run E@H and there are times when both are running on the same GPU just fine.

Peter Hucker
Peter Hucker
Joined: 12 Aug 06
Posts: 244
Credit: 218,670,728
RAC: 682,306

Does one slow the other

Does one slow the other down?  If not you should run them both all the time.

mmonnin
mmonnin
Joined: 29 May 16
Posts: 265
Credit: 924,144,806
RAC: 275,056

E@H only runs when MW is out

E@H only runs when MW is out of work because of its stupid 10min forced wait. Of course they share resources for a bit but I don't run like that very long.

Peter Hucker
Peter Hucker
Joined: 12 Aug 06
Posts: 244
Credit: 218,670,728
RAC: 682,306

MW needs to get its act

MW needs to get its act together.  No other project restricts how much work you can get.  You'd think they'd want us to process as much as possible.  If I can't get work from a project, I'm going to go elsewhere.

Comment viewing options

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