Folding Forum

Devildoll wrote:or if nvidia cards/openCL implementation has a predefined edge on amd cards when it comes to this kind of compute work, if so , it would be cool to know what attributes of the cards govern their ability to do this kind of work.


But then , points arn't the most important thing here , getting the WU's done is the main thing , but the points/competition aspect of F@H still interests me.

It could be any of the factors you've mentioned.

Also, we're comparing two different things. Core 16 uses OpenCL on AMD. Core 15 uses CUDA on NV. Nobody has said anything about OpenCL on NV yet. There's no urgent need to build such a core since the CUDA version is doing fine, but all three would be OpenMM so it does make sense to consider all three eventually.

I was about to try the new core, but when it starts I get the following error, repeated 3 times and then the client stops trying.

Code: Select all

14:59:37:Received Unit: id:02 state:DOWNLOAD project:11293 run:31 clone:93 gen:0 core:0x16 unit:0x000000006652edbc4d94bb49918ad121
14:59:37:Starting Unit 02
14:59:37:Running core: C:/Users/Fredrik/AppData/Roaming/FAHClient/cores/www.stanford.edu/~pande/Win32/x86/ATI/R600/Core_16.fah/FahCore_16.exe -dir 02 -suffix 01 -lifeline 5116 -version 701 -checkpoint 15 -gpu 0
14:59:37:Started core on PID 6072
14:59:37:FahCore 0x16 started
14:59:37:Unit 02:
14:59:37:Unit 02:*------------------------------*
14:59:37:Unit 02:Folding@Home GPU Core
14:59:37:Unit 02:Version 2.11 (Thu Dec 9 15:00:14 PST 2010)
14:59:37:Unit 02:
14:59:37:Unit 02:Compiler  : Microsoft (R) 32-bit C/C++ Optimizing Compiler Version 15.00.30729.01 for 80x86
14:59:37:Unit 02:Build host: user-f6d030f24f
14:59:37:Unit 02:Board Type: AMD/OpenCL
14:59:37:Unit 02:Core      : x=16
14:59:37:Unit 02: Window's signal control handler registered.
14:59:37:Unit 02:Preparing to commence simulation
14:59:37:Unit 02:- Looking at optimizations...
14:59:37:Unit 02:DeleteFrameFiles: successfully deleted file=02/wudata_01.ckp
14:59:37:Unit 02:- Created dyn
14:59:37:Unit 02:- Files status OK
14:59:37:Unit 02:sizeof(CORE_PACKET_HDR) = 512 file=<>
14:59:37:Unit 02:- Expanded 42526 -> 171163 (decompressed 402.4 percent)
14:59:37:Unit 02:Called DecompressByteArray: compressed_data_size=42526 data_size=171163, decompressed_data_size=171163 diff=0
14:59:37:Unit 02:- Digital signature verified
14:59:37:Unit 02:
14:59:37:Unit 02:Project: 11293 (Run 31, Clone 93, Gen 0)
14:59:37:Unit 02:
14:59:37:Unit 02:Assembly optimizations on if available.
14:59:37:Unit 02:Entering M.D.
14:59:39:Unit 02:Tpr hash 02/wudata_01.tpr:  2248307142 147158541 3598004666 116545269 1841948847
14:59:39:Unit 02:Working on ALZHEIMER DISEASE AMYLOID
14:59:39:Unit 02:Client config unavailable.
14:59:39:Unit 02:Starting GUI Server
14:59:41:Unit 02:Finished fah_main
14:59:41:Unit 02:
14:59:41:Unit 02:Successful run
14:59:41:Unit 02:DynamicWrapper: Finished Work Unit: sleep=10000
14:59:51:Unit 02:Reserved 0 bytes for xtc file; Cosm status=0
14:59:51:Unit 02:Reserved 0 0 786430464 bytes for arc file=<02/wudata_01.trr> Cosm status=0
14:59:51:Unit 02:Allocated 0 bytes for edr file
14:59:51:Unit 02:Error: could not open bedfile, but going on anyway
14:59:51:Unit 02:- Checksum of file (02/wudata_01.edr) read from disk doesn't match
14:59:51:Unit 02:edrfile file hash check failed.
14:59:51:Unit 02:
14:59:51:Unit 02:Folding@home Core Shutdown: FILE_IO_ERROR
14:59:52:FahCore running Unit 02 returned: FILE_IO_ERROR (117)
14:59:52:WARNING: Unit 02 Fatal error, dumping
14:59:52:Sending unit results: id:02 state:SEND project:11293 run:31 clone:93 gen:0 core:0x16 unit:0x000000006652edbc4d94bb49918ad121
14:59:52:Connecting to 171.67.108.44:8080
14:59:53:Server responded WORK_ACK (400)
14:59:53:Cleaning up Unit 02

Just a little FYI..

I'm running 4 clients on 2 5870 cards.

http://www.overclock.net/overclock-net-folding-home-team/983961-v7-core16-4-clients-2-cards.html

did you just add two slots onto one gpu ? or did you have to configure more stuff than in normal cases.

That looks interesting. Mind you, you need decent amount of ram. And what about cpu cores? Are you giving each client its core? Or just 1 per card? Wonder what would 6970 would produce with its 2gb of ram.
until bigger wus come our way it might be good way to use our high end cards

grunion wrote:Just a little FYI..

I'm running 4 clients on 2 5870 cards.

http://www.overclock.net/overclock-net-folding-home-team/983961-v7-core16-4-clients-2-cards.html

I can confirm this works. Infact, there is something interesting that I found when running two instances. Right now, I am running a core 11 and core 16 on HD 5870 and in an unexpected twist, the cpu time they use is considerably lower. I was seeing cpu time of maybe 20% for core 16 alone. core 11 tend to normally be 25% cpu (on quad core cpu btw). Now fah 16 is about 13% cpu and core 11 about 6%. this makes me a happier camper.

when running two instances, the TPF on core 16 is roughly the same as if it was the only gpu core running. The penalty is only a few seconds per frame on core 16, to run two instances.

Working fine for me on a HD 5870, getting ~8500 PPD so far :-



I'm also running the SMP on a i7-2600K @ 4.6Ghz, so I've just left the card running a single instance.

your SMP client must be suffering a lot, as that GPU client eats at least one core, sometimes jumping into second core.. So it is advisable to do smp 7 with the cpu and leave one thread to GPU.

muziqaz wrote:your SMP client must be suffering a lot, as that GPU client eats at least one core, sometimes jumping into second core.. So it is advisable to do smp 7 with the cpu and leave one thread to GPU.

Yep fair point, I've knocked a thread off the SMP and now the HD 5870 has improved a bit :-



Even with SMP 7, the 2600K is still comfortably over 20K PPD on it's own

I'm going to try that multiple client trick. Under the right circumstances it might defeat the time wasted in spin-waits. One thing I've found it that if the GPU slot and the SMP slot run at the same priority, the GPU does slow down some, but the wasted CPU time goes way down. Eventually they'll be doing some optimiation that will probably help, but in the meantime, any tricks you discover are worth trying even if they become obsolete soon.

Any idea whether multiple clients are worthwhile for a 5850?