Folding Forum

k1wi wrote:I have started a new thread where people can post their results:

Might keep this thread clean for debugging/feedback purposes.

viewtopic.php?f=38&t=23440

Good idea, k1wi.

@proteneer
The benchmark provides the type of general results that a benchmark should provide but FAH seems to really only need Single. Do you recommend that k1wi keep both single and double? Do you recommend that he keep both CUDA and OpenCL? ... and more importantly than simple Y/N responses, why?

nice - its a nice list of numbers the you guys are getting. the numbers are going to vary in the upcoming months, as we begin to heavily optimize OpenMM. I trust all of you guys here, but you might want to mention that those numbers are on unofficial - it's pretty easy to hack so we keep an official list on fahbench.com as well.

You can keep both the CUDA/OpenCL numbers, we care more about single precision perf. at this point. Don't really need double (though it is one area in which the K20 shines).

PS - does anyone have overclocked 580? That card is a BEAST.

proteneer wrote:nice - its a nice list of numbers the you guys are getting. the numbers are going to vary in the upcoming months, as we begin to heavily optimize OpenMM. I trust all of you guys here, but you might want to mention that those numbers are on unofficial - it's pretty easy to hack so we keep an official list on fahbench.com as well.

You can keep both the CUDA/OpenCL numbers, we care more about single precision perf. at this point. Don't really need double (though it is one area in which the K20 shines).

PS - does anyone have overclocked 580? That card is a BEAST.

I have updated the page to note that it is unofficial and dependent on the integrity of the folder. Seeing as anyone can submit their value any erranous values should stand out like a sore thumb.

You'll need to wait until Grandpa pops back in and turns his 580 over at 1GHz+, then you'll see a beast of a result...

[Add]At some point would it be possible to add the version number to the splash ASCII art? About where it says Beta...

Thanks - I will add in a version number once we're stable (there are too many revisions right now to make a version number meaningful)

@ k1wi, you can add the results of Tesla K20 and GTX 660Ti from http://fahbench.com.
And add in the text, those who already have results is desirable to test the next versions of "FAHBench" on the same frequencies of GPU (MEM). In order to compare the results correctly.

How about adding clock rate and lspci codes (which can be traced back to the real hardware) to the output. We can ask the donors to report how much they're overclocking but that becomes hit-and-miss so a direct report, if you can provide it, would be useful to those of us reading the forum, whether you need that in the information you're gathering for FAH or not.

Since you're reporting the results in ns/day I'm guessing that you plan to keep the same protein(s) "forever" so knowing the number of atoms and the number of steps will never be needed, but if there's a chance you'd need to try a new protein, that might be useful reference information.

I ran into that problem too.

The benchmark does not currently handle mixed environments for OpenCL (i.e. AMD & NVidia in the same system). Do/did you have an AMD GPU or CPU?

k1wi wrote:I ran into that problem too.

The benchmark does not currently handle mixed environments for OpenCL (i.e. AMD & NVidia in the same system). Do/did you have an AMD GPU or CPU?

AMD/ATI GPU is not set to fold and seems to be DeviceId 0, Nvidia GTX 570 seems to be DeviceId 1. Nothing was folding at the time of the test.

Any insight on the --display-devices failure on the Nvidia, but ID'd the AMD/ATI??

PinHead wrote:AMD/ATI GPU is not set to fold and seems to be DeviceId 0, Nvidia GTX 570 seems to be DeviceId 1. Nothing was folding at the time of the test.

Any insight on the --display-devices failure on the Nvidia, but ID'd the AMD/ATI??

Yeah, I don't even have my ol HD4770 in my rig at the moment and I get the same issue (except that it doesn't even get as far as you do on --display-devices).

proteneer would be able to give you a better explanation as to why the NVidia device fails. It could be as simple as it being the second device.

As I understand it support for heterogenous drivers/cards will be added in a future, hopefully next, release. In the mean time we just have to wait and play with the CUDA stats .

BTW: He added the following to the fahbench.com page:

Latest NVIDIA/ATI Drivers (Exclusive-Or)

As I understand it support for heterogeneous drivers/cards will be added in a future, hopefully next, release. In the mean time we just have to wait and play with the CUDA stats .

BTW: He added the following to the fahbench.com page:

Latest NVIDIA/ATI Drivers (Exclusive-Or)

I previously have folded on both AMD and NVidia GPUs (exclusive "and" )in this system. I didn't uninstall the other drivers (yet) but it has been convenient to just unplug one GPU and plug in the other. The Windows Device Manager shows only the current hardware but the previous drivers are still installed. I'm not sure yet if that matters or not.

Also on my list of "I wonder if...", I have a remote system which is folding on a recent NV GPU but my display is attached to legacy ATI device (unsupported).

It should be noted that there are numerous FAH Donors with mixed GPU systems and support for this environment has been requested many, many times. Some donors made it work in V6 in spite if the difficulties and V7 has made it somewhat easier but still not a slam-dunk. There have been a number of V7 tickets associated with support of the mixed environment, including a couple which are still open.

The monochromatic system may be relatively common for OpenMM's stand-alone academic customers, but FAH Donors want more.

I have an issue where the CUDA benchmarks run fine but the OpenCL ones do not. After a few seconds, I get a standard Windows error message telling me FAHBench.exe has crashed, but there's no text output to the command line. I have installed VS2008 through the link on fahbench.com, and I made sure to install vcsetup instead of one of the other three (I did accidentally run vcssetup as well, but I don't think having two of them installed is a problem). The PATH environment variable was copied directly from the site, and it's set as both a System and User variable. Here's the Windows crash log output if it's any help:
And on a somewhat related note, watch out for Windows trying to roll back your CUDA version. I have the 310.90 drivers installed, but I think Windows reverted back to version 4.something during one of my reboots since the CUDA tests didn't run any more after that "update" with that 126 error indicative of CUDA 5 being missing. Re-installing the CUDA 5 production release got it working again.

PinHead wrote:It doesn't seem to like Nvidia/Nvidia either:

Vista Ultimate 32 bit, Tesla C2050 and GTX550Ti

FAHBench.exe --display-devices
Output:

Code: Select all

                                                                               
                                          O              O                     
   P R O T E N E E R     C--N              \              \               N    
                         |                  C              C=O           / \-C 
                         C                 /               |          N-C     \
  .C-C                 C/                  C               C           |      C
 /    \          O     |                   |               /           N      |
C     C          |     |           O       C              C                 /-C
 \_N_/ \   N    _C_    C           |      /         O    /                 C   
        C-/ \_C/   \N-/ \    N   /-C-\   C          |    |           O    /    
        |     |           C-/ \C/     N-/ \_   N\  /C\  -C      N    |    |    
        O     |           |    |            \C/  C/   N/  \_C__/ \   C-\  C    
              C           O    |             |   |          |     C-/   N/ \-C
               \_C             C             O   |          O     |          | 
                  \             \-O              C                C          O 
                  |                               \                \           
                  C    N         Folding@Home      C--N             C          
                   \   |      Benchmark  (Beta)    |                |          
                    N--C                           O                |          
                        \        Yutong Zhao                       C=O        
                         N    proteneer@gmail.com                 /           
                                                                 O            
                                                                               
               for official stats, please visit http://www.fahbench.com               

=== 1 OpenCL platform(s) found: ===
  -- 0 --
  PROFILE = FULL_PROFILE
  VERSION = OpenCL 1.1 CUDA 4.2.1
  NAME = NVIDIA CUDA
  VENDOR = NVIDIA Corporation
  EXTENSIONS = cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_d3d9_sharing cl_nv_d3d10_sharing cl_khr_d3d10_sharing cl_nv_d3d11_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll 
=== 1 OpenCL device(s) found on platform:
  -- 0 --
  DEVICE_NAME = Tesla C2050 / C2070
  DEVICE_VENDOR = NVIDIA Corporation
  DEVICE_VERSION = OpenCL 1.1 CUDA
  DRIVER_VERSION = 306.94
  DEVICE_MAX_COMPUTE_UNITS = 14
  DEVICE_MAX_CLOCK_FREQUENCY = 1147
  DEVICE_GLOBAL_MEM_SIZE = 2818572288
  -- 1 --
  DEVICE_NAME = GeForce GTX 550 Ti
  DEVICE_VENDOR = NVIDIA Corporation
  DEVICE_VERSION = OpenCL 1.1 CUDA
  DRIVER_VERSION = 306.94
  DEVICE_MAX_COMPUTE_UNITS = 4
  DEVICE_MAX_CLOCK_FREQUENCY = 1962
  DEVICE_GLOBAL_MEM_SIZE = 1073741824
Invalid Platform (please use either OpenCL or CUDA)

Tesla C2050 WDDM mode ( OpenCL )
FAHBench.exe -deviceId 0 -platform OpenCL -precision single
Explicit:
13.5521 ns/day ( accuracy really beats on the slow drive, factor ? )

Implicit:
61.4903 ns/day

GTX550Ti ( OpenCL )FAHBench.exe -deviceId 1 -platform OpenCL -precision single
Explicit:
8.40868 ns/day

Implicit:
36.6661 ns/day

Tesla C2050 WDDM mode ( CUDA )FAHBench.exe -deviceId 0 -platform CUDA -precision single
Explicit:
15.8791 ns/day

Implicit:
55.0354 ns/day

GTX550Ti ( CUDA )FAHBench.exe -deviceId 1 -platform CUDA -precision single
Explicit:
Error launching CUDA compiler:-1
nvcc : fatal error : Value 'compute_21' is not defined for option
'gpu-architecture'

Is this with Version 0_2?

I have a pretty good idea why the NVIDIA/ATI hybrid is crashing - will be fixing it soon.

Fixed mixed ATI/NVIDIA bug

Some side questions/observations:

I monitor the %GPU usage on the systray and noticed that most of the time running the GPU %Used stays at 0%. So what is it actually doing with all that CPU time?

What is being checked for accuracy? I observed that the %GPU usage does not go above 0% till after the first checking for accuracy...done.