GPU use ??

Message boards : Number crunching : GPU use ??

To post messages, you must log in.

AuthorMessage
White_Bear

Send message
Joined: 24 Jan 11
Posts: 1
Credit: 11,504
RAC: 0
Message 69483 - Posted: 25 Jan 2011, 7:26:50 UTC

Sorry I'm new here.

Am I correct in thinking that R@H does not use GPUs ??

Is that because of the type(s) of calculations being done ??
ID: 69483 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
mikey
Avatar

Send message
Joined: 5 Jan 06
Posts: 1894
Credit: 8,555,637
RAC: 10,862
Message 69484 - Posted: 25 Jan 2011, 10:49:23 UTC - in response to Message 69483.  

Sorry I'm new here.

Am I correct in thinking that R@H does not use GPUs ??

Is that because of the type(s) of calculations being done ??


yes you are correct that Rosetta does not have a gpu app. It is because the current app is waaaaay too big to fit onto a gpu, it can easily take more than 2 gig of memory, and if you start swapping memory on a gpu all efficiency is lost. They have tested it and said when/if it becomes a viable alternative they will introduce one.
ID: 69484 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Dirk Broer

Send message
Joined: 16 Nov 05
Posts: 22
Credit: 2,812,128
RAC: 10,239
Message 69693 - Posted: 23 Feb 2011, 0:08:47 UTC - in response to Message 69484.  
Last modified: 23 Feb 2011, 0:09:33 UTC

yes you are correct that Rosetta does not have a gpu app. It is because the current app is waaaaay too big to fit onto a gpu, it can easily take more than 2 gig of memory, and if you start swapping memory on a gpu all efficiency is lost. They have tested it and said when/if it becomes a viable alternative they will introduce one.


Is that true? How can it be then that I can run Rosetta on machines with only 1 Gb of memory? Isn't it true that there are GPUs on the market with up to 6 Gb of memory (nVidia Quadro 6000, Tesla stations). You can also make a smaller application for the GPU to use it efficiently.
ID: 69693 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Dragokatzov

Send message
Joined: 5 Oct 05
Posts: 25
Credit: 2,446,376
RAC: 0
Message 69696 - Posted: 23 Feb 2011, 11:48:52 UTC

tesla and quadro do have lots of ram. but they are also very pricy. however, they are very powerful beasts, and would provide a lot more flops. perhaps looking into a different area of research, and creating a new app for that requiring less ram...
Victory is the ONLY option!
ID: 69696 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
mikey
Avatar

Send message
Joined: 5 Jan 06
Posts: 1894
Credit: 8,555,637
RAC: 10,862
Message 69697 - Posted: 23 Feb 2011, 13:17:03 UTC - in response to Message 69693.  

yes you are correct that Rosetta does not have a gpu app. It is because the current app is waaaaay too big to fit onto a gpu, it can easily take more than 2 gig of memory, and if you start swapping memory on a gpu all efficiency is lost. They have tested it and said when/if it becomes a viable alternative they will introduce one.


Is that true? How can it be then that I can run Rosetta on machines with only 1 Gb of memory?

Because your cpu will use the hard drive as backup for any memory shortfalls it runs into. If your gpu did that it would negate all the advantages of using the gpu in the first place.

Isn't it true that there are GPUs on the market with up to 6 Gb of memory (nVidia Quadro 6000, Tesla stations).

Yes there are and here is one for sale:
PNY VCQ6000-PB Quadro 6000 6GB 384-bit GDDR5 PCI Express 2.0 x16 Workstation Video Card PNY VCQ6000-PB Quadro 6000 6GB 384-bit GDDR5 PCI Express 2.0 x16 Workstation Video Card

* Quadro 6000
* 6GB GDDR5
* PCI Express 2.0 x16

* Chipset Manufacturer: NVIDIA
* Stream Processors: 448
* DirectX: DirectX 11
* OpenGL: OpenGL 4.0
* Model #: VCQ6000-PB
* Item #: N82E16814133347
* Return Policy: VGA Standard Return Policy

* Was: $3,999.99
* Now: $3,799.99
Not EXACTLY cost effective! 448 stream processors for 3800 US Dollars!!! You can get an ATI 5770 with 800 stream processing units for about 100 US Dollars after rebate! And you can get an Nvidia 470 with 448 stream processing units for about 230 US Dollars after rebate! Both the ATI 5770 and the Nvidia 470 have around 1 gigabyte of memory on them.

You can also make a smaller application for the GPU to use it efficiently.

That was the point of my post, yes you can, BUT it won't do as much work in the time that a cpu can! You would have to do 3 or 4 times as many units to do the same amount of work as the cpu units do now, and there goes your time savings by using the gpu!

They REALLY HAVE thought this out and we REALLY ARE doing it the best way for now.
ID: 69697 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Dirk Broer

Send message
Joined: 16 Nov 05
Posts: 22
Credit: 2,812,128
RAC: 10,239
Message 69699 - Posted: 23 Feb 2011, 22:11:41 UTC - in response to Message 69697.  
Last modified: 23 Feb 2011, 22:30:27 UTC

You can also make a smaller application for the GPU to use it efficiently.

That was the point of my post, yes you can, BUT it won't do as much work in the time that a cpu can! You would have to do 3 or 4 times as many units to do the same amount of work as the cpu units do now, and there goes your time savings by using the gpu!

They REALLY HAVE thought this out and we REALLY ARE doing it the best way for now.


There are lots of people with a HD 5970, which offers 5 TFlop of computing power, and crunches WUs up to 20-25 times faster than CPU applications. The new HD 6990 wil even offer 6 TFlop, and I can get a HD 5970 for a mere 350 Euro, so how about cost effective? The HD 6990 will cost around 500 Euro and will also attract BOINC entusiasts. If you look into other project statistics you will encounter Fermi-based Tesla stations and real heavy Quadro's. Some people/institutes really can afford them, why not make use of them?

P.S. A HD 5970 has 3200 Stream processors and 2 Gb of RAM. The HD 6990 will have 3820 stream processors, and 4 Gb of RAM.
ID: 69699 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Murasaki
Avatar

Send message
Joined: 20 Apr 06
Posts: 303
Credit: 511,418
RAC: 0
Message 69701 - Posted: 24 Feb 2011, 1:32:19 UTC - in response to Message 69693.  
Last modified: 24 Feb 2011, 1:51:52 UTC

yes you are correct that Rosetta does not have a gpu app. It is because the current app is waaaaay too big to fit onto a gpu, it can easily take more than 2 gig of memory, and if you start swapping memory on a gpu all efficiency is lost. They have tested it and said when/if it becomes a viable alternative they will introduce one.


Is that true? How can it be then that I can run Rosetta on machines with only 1 Gb of memory?


It is a slight exaggeration. Larger proteins take more memory to process than shorter proteins, so you will see the memory requirement vary between tasks. However I have never seen claims of memory use above 1.5GB per task except when there was a bug in the Rosetta code.

Many Rosetta tasks use less than 500MB.
ID: 69701 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile robertmiles

Send message
Joined: 16 Jun 08
Posts: 1223
Credit: 13,824,497
RAC: 2,340
Message 69807 - Posted: 13 Mar 2011, 5:02:16 UTC - in response to Message 69699.  

P.S. A HD 5970 has 3200 Stream processors and 2 Gb of RAM. The HD 6990 will have 3820 stream processors, and 4 Gb of RAM.


Without a major rewrite of the program, you can actually use only as many of those stream processors as you can find around 500 MB of memory each for.

Therefore, an HD 6990 should give a speedup approximately 8 times as fast as a CPU workunit. Other than the 8 stream processors actually running the calculations, the other stream processors might be used a little for accessing memory that the 8 cannot reach directly, but not for much more.
ID: 69807 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote

Message boards : Number crunching : GPU use ??



©2024 University of Washington
https://www.bakerlab.org