Folding@Home Benchmark 1.2.0 (GUI)

[Napoleon]

http://fahbench.com

Latest 0_3

Feel free to give it a whirl and let us know what you think.

The download link seems to be outdated. I managed to pull 0_3 from http://proteneer.com/fahbench/FAHBench_0_3.zip. Definitely giving it a whirl. BTW, I figured I won't bother benchmarking my ION anymore. I just use it to drive my display and occasionally for Einstein@Home, these days.

Compute Capability 1.1 (no DP support, limited support for CUDA atomic operations, only OpenCL 1.0) very effectively limits it to OpenCL SP benchmarking, nothing more.

[proteneer]

We only need to use a small % of time on your GPU to figure out its folding capabilities. The rest of the time is spent on various overhead like checking for accuracy (against the CPU version), setting up the system, creating the contexts, etc.

PS. Updated download link - thanks Napolean

[Jesse_V]

I was fiddling with a modified version of P5-133XL's batch file, and I got a couple of warnings about platformId not being specified:

Code: Select all

C:\Users\Admin\Downloads\bench>bench

C:\Users\Admin\Downloads\bench>fahbench -deviceId 0 -platform CUDA -precision si
ngle

                                          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 www.fahbench.com

Explicit:
Checking for accuracy...done
18.0267 ns/day
Implicit:
Checking for accuracy...done
67.7063 ns/day

C:\Users\Admin\Downloads\bench>fahbench -deviceId 0 -platform CUDA -precision si
ngle --disable-splash
Explicit:
Checking for accuracy...done
18.0587 ns/day
Implicit:
Checking for accuracy...done
67.5475 ns/day

C:\Users\Admin\Downloads\bench>fahbench -deviceId 0 -platform OpenCL -precision
single --disable-splash
Warning: Using OpenCL platform but no platformId specified, setting platformId=0

Explicit:
Checking for accuracy...done
9.18338 ns/day
Implicit:
Checking for accuracy...done
63.9526 ns/day

C:\Users\Admin\Downloads\bench>fahbench -deviceId 0 -platform OpenCL -precision
single --disable-splash
Warning: Using OpenCL platform but no platformId specified, setting platformId=0

Explicit:
Checking for accuracy...done
9.75621 ns/day
Implicit:
Checking for accuracy...done
67.3422 ns/day

C:\Users\Admin\Downloads\bench>

This is on version 0.3. (I've already reported my numbers with v0.2 on the other thread)

[proteneer]

Yep those warnings are normal. OpenCL is finnicky in that it can detect multiple platforms (--display-devices) to see them all. This way you can now specify both the platform and the particular device.

[Napoleon]

After all I decided to benchmark my ION one last time, for the sake of consistency. Now that I've jumped through the hoops to get the benchmark to run, I'll be reporting results in the Most Powerful GPU (FAHBench) thread. Not that I have the slightest chance of clocking top ns/day results, but it gives me the warm fuzzy feeling of actually being... a beta tester.

[k1wi]

Confirmed that v0.3 works with AMD & NVidia drivers (with only an NVidia card installed). Setting platformId was straightforward.

I'm not sure about my OpenCL results though as I had a really high implicit ns/day result on my last OpenCL test.

The numbers at the top of the screenshot are from a previous test 18MHz lower, but somehow the extra 18MHz was good for 16 ns/day! (http://i49.tinypic.com/35b5xef.png). I just repeated the test two more times and got the same 114 ns/day implicit on OpenCL.

So not quite sure why that would be. The results are higher than the CUDA result for that GPU, which seems to be unlikely given the trends observed, unless there has been an OpenCL improvement with v3?

[proteneer]

Confirmed that v0.3 works with AMD & NVidia drivers (with only an NVidia card installed). Setting platformId was straightforward.

I'm not sure about my OpenCL results though as I had a really high implicit ns/day result on my last OpenCL test.

The numbers at the top of the screenshot are from a previous test 18MHz lower, but somehow the extra 18MHz was good for 16 ns/day! (http://i49.tinypic.com/35b5xef.png). I just repeated the test two more times and got the same 114 ns/day implicit on OpenCL.

So not quite sure why that would be. The results are higher than the CUDA result for that GPU, which seems to be unlikely given the trends observed, unless there has been an OpenCL improvement with v3?

Do you mind showing the --display-devices ?

[k1wi]

I'm accessing the computer at home from work via LogMeIn (if that makes any difference) so can only take screen grabs:
http://i50.tinypic.com/2vwuycj.png

From how I understand explicit calculations on the GPU are performed in conjunction with the CPU, does the benchmark depend on the single thread performance of the CPU? Or are the calculations purely GPU based?

[bruce]

I'm accessing the computer at home from work via LogMeIn (if that makes any difference) so can only take screen grabs:
http://i50.tinypic.com/2vwuycj.png

From how I understand explicit calculations on the GPU are performed in conjunction with the CPU, does the benchmark depend on the single thread performance of the CPU? Or are the calculations purely GPU based?

I would never use the word "purely" in that context. At the very least, the CPU needs to arrange work into some kind of packets to be sent to the GPU and receive results. Since CPUs are good at atomic functions, the CPU may also be used for limited analysis of the results, such as precision verification. When a GPU processes real graphics, a lot of work has been done in analyzing how much time is spent with the CPU, with the PCIe bus, with the shaders, and with other things like Anti-Aliasing, etc. I'm not sure how much of that type of optimization has been done for GROMACS but it can probably benefit greatly from the same kind of analysis since the atoms form a "sparse matrix" of data. From the questions proteneer is proposing on his home-page, it's obvious to me that he's going to make great strides in those areas.

Balancing the work best done by a GPU and by a CPU working in tandem, even given a very wide range of relative speed variations would be a very simplistic first step toward optimization, including avoiding making the GPU wait for the CPU unnecessarily when folding or when actually timing the benchmark.

[PinHead]

FAHBench 0.3
Win 7 64bit
AMD and NVidia card installed

FAHBench --display-devices
output:

Code: Select all

[2] compatible platform(s):
  -- 0 --
  PROFILE = FULL_PROFILE
  VERSION = OpenCL 1.1 AMD-APP (831.4)
  NAME = AMD Accelerated Parallel Processing
  VENDOR = Advanced Micro Devices, Inc.
  -- 1 --
  PROFILE = FULL_PROFILE
  VERSION = OpenCL 1.1 CUDA 4.2.1
  NAME = NVIDIA CUDA
  VENDOR = NVIDIA Corporation

(2) device(s) found on platform 0:
  -- 0 --
  DEVICE_NAME = Juniper
  DEVICE_VENDOR = Advanced Micro Devices, Inc.
  DEVICE_VERSION = OpenCL 1.1 AMD-APP (831.4)

  -- 1 --
  DEVICE_NAME = Intel(R) Core(TM)2 Quad CPU    Q9550  @ 2.83GHz
  DEVICE_VENDOR = GenuineIntel
  DEVICE_VERSION = OpenCL 1.1 AMD-APP (831.4)

(1) device(s) found on platform 1:
  -- 0 --
  DEVICE_NAME = GeForce GTX 570
  DEVICE_VENDOR = NVIDIA Corporation
  DEVICE_VERSION = OpenCL 1.1 CUDA

Invalid Platform (please use either OpenCL or CUDA)

I don't think I have been running FAHBench correctly for having 2 cards.

If I ( now ) understand the output above, then:

FAHBench.exe -deviceId 0 -platform OpenCL -precision single ( would run on the AMD GPU )
FAHBench.exe -deviceId 1 -platform OpenCL -precision single ( would run on the CPU )
FAHBench.exe -deviceId 0 -platform CUDA -precision single ( would run on the GTX570 )

No option for testing OpenCL on the GTX570 even though it ID's the device as OpenCL 1.1 ?

Am I reading and interpreting this information correctly for the command line arguements?

[k1wi]

I was perhaps too loose with my use of 'purely'.

My question arose from an observation that the benchmark app uses an entire CPU 'thread' during the actual speed calculation side of the measurements. As I get into 1GHz+ it looks like GPU utilisation may be dropping, perhaps because it is being bottlenecked by the CPU. This may be an issue for people with CPUs that have less powerful single thread CPUs, or more accurately, may mean that the CPU being used may need to be taken into account.

Aside from 'housekeeping' (for want of a better term) services provided by the CPU such as you noted, I was wondering whether the test is offloading some protein folding calculations (i.e. non-bonded force calculations) to the GPU (GPU acceleration), or whether they were performing all the calculations on the GPU (and only using the CPU for housekeeping and feeding the GPU).

[bruce]

I've heard some suggestions about partitioning work between a CPU and a GPU in different ways than is now being done. It would seem like a real challenge to decide how much work to assign to each one -- (see my earlier mention load balancing for the really wide range of relative speeds).

[art_l_j_PlanetAMD64]

I've heard some suggestions about partitioning work between a CPU and a GPU in different ways than is now being done. It would seem like a real challenge to decide how much work to assign to each one -- (see my earlier mention load balancing for the really wide range of relative speeds).

Yes, depending on how much more CPU resources were used by 'GPU folding', NVidia owners might have to allocate a core or two away from SMP folding, like AMD/ATI owners do now, especially in multi-GPU systems.

[k1wi]

FAHBench 0.3
Win 7 64bit
AMD and NVidia card installed

FAHBench --display-devices
output:

Code: Select all

[2] compatible platform(s):
  -- 0 --
  PROFILE = FULL_PROFILE
  VERSION = OpenCL 1.1 AMD-APP (831.4)
  NAME = AMD Accelerated Parallel Processing
  VENDOR = Advanced Micro Devices, Inc.
  -- 1 --
  PROFILE = FULL_PROFILE
  VERSION = OpenCL 1.1 CUDA 4.2.1
  NAME = NVIDIA CUDA
  VENDOR = NVIDIA Corporation

(2) device(s) found on platform 0:
  -- 0 --
  DEVICE_NAME = Juniper
  DEVICE_VENDOR = Advanced Micro Devices, Inc.
  DEVICE_VERSION = OpenCL 1.1 AMD-APP (831.4)

  -- 1 --
  DEVICE_NAME = Intel(R) Core(TM)2 Quad CPU    Q9550  @ 2.83GHz
  DEVICE_VENDOR = GenuineIntel
  DEVICE_VERSION = OpenCL 1.1 AMD-APP (831.4)

(1) device(s) found on platform 1:
  -- 0 --
  DEVICE_NAME = GeForce GTX 570
  DEVICE_VENDOR = NVIDIA Corporation
  DEVICE_VERSION = OpenCL 1.1 CUDA

Invalid Platform (please use either OpenCL or CUDA)

I don't think I have been running FAHBench correctly for having 2 cards.

If I ( now ) understand the output above, then:

FAHBench.exe -deviceId 0 -platform OpenCL -precision single ( would run on the AMD GPU )
FAHBench.exe -deviceId 1 -platform OpenCL -precision single ( would run on the CPU )
FAHBench.exe -deviceId 0 -platform CUDA -precision single ( would run on the GTX570 )

No option for testing OpenCL on the GTX570 even though it ID's the device as OpenCL 1.1 ?

Am I reading and interpreting this information correctly for the command line arguements?

Hi PinHead,

I've rearranged the flags so that it is a little more logical
FAHBench.exe -platform OpenCL -platformId 0 -deviceId 0 -precision single ( would run on the AMD GPU )
FAHBench.exe -platform OpenCL -platformId 0 -deviceId 1 -precision single ( would run on the CPU )
FAHBench.exe -platform OpenCL -platformId 1 -deviceId 0 -precision single ( would run on the GTX570 )

Where:
-platform OpenCL signals that we're using OpenCL
-platformId separates between the AMD and NVidia implemetation (I guess this is technically the 'driver')
-deviceId distinguishes between the devices on each platform.

Add: it was the "setting platformId=0" in this 'warning' that made me work out the platformId bit:

Code: Select all

Warning: Using OpenCL platform but no platformId specified, setting platformId=0

[PinHead]

K1wi,

So -platform will relate to OpenCL or CUDA and has nothing to do with the platformId?
And platformId is just the order in which the cards were detected in ( not sure why the CPU was detected ) ?

Getting the "Warning: Using OpenCL platform but no platformId specified, setting platformId=0" on both machines with v0.3, don't think I received it on v0.1.