【发现】core21内置性能监测功能

wpf999

horst1981 发表于 2016-2-12 22:22
但是数值听起来PPD更屌啊
我ns/day七点几几,和我PPD二十几万，一般观感来讲后者听起来好像更牛逼[e ...

从并行计算原理上讲980的ns/秒性能不可能超过960的2倍，但是980的ppd性能超过960的2倍。

wpf999

wpf999 发表于 2016-2-5 01:09
Project: 11401 (Run 0, Clone 15, Gen 86)
GTX970
蛋白质折叠性能：37.8ns/day

log.txt (7.93 KB, 下载次数: 1020)

2016-2-12 22:50 上传

点击文件名下载附件

Project: 11401 (Run 11, Clone 5, Gen 78)
GTX960
蛋白质折叠性能：24.5ns/day


970/960性能比=37.8/24.5=1.54

970/960CUDA核数比=1664/1024=1.625

wpf999

horst1981 发表于 2016-2-12 22:22
但是数值听起来PPD更屌啊
我ns/day七点几几,和我PPD二十几万，一般观感来讲后者听起来好像更牛逼[e ...

请见17楼数据分析。980也会得出类似结论

金鹏

wpf999 发表于 2016-2-12 22:29
从并行计算原理上讲980的ns/秒性能不可能超过960的2倍，但是980的ppd性能超过960的2倍。 ...

因为有QRB的非线性,所以性能差35%+就会导致PPD翻倍

coju

A卡为啥没有？

coju

Project: 10468 (Run 0, Clone 494, Gen 188)

Unit: 0x0000013a538b3db9538cb7025b5b757d

CPU: 0x00000000000000000000000000000000

Machine: 1

Reading tar file state.xml

Reading tar file system.xml

Reading tar file integrator.xml

Reading tar file core.xml

Digital signatures verified

**************************** Zeta Folding@home Core ****************************
       Type: 23
       Core: Zeta
    Website: http://folding.stanford.edu/
  Copyright: (c) 2009-2014 Stanford University
     Author: Yutong Zhao <[email protected]>
       Args: -dir 01 -suffix 01 -version 704 -lifeline 8444 -checkpoint 15 -gpu
             0 -gpu-vendor ati
     Config: <none>
************************************ Build *************************************
    Version: 0.0.55
       Date: Mar 27 2014
       Time: 15:06:29
    SVN Rev: Unknown
     Branch: Unknown
   Compiler: Visual C++ 2008
    Options: $( /TP $) /nologo /EHa /wd4297 /wd4103 /Ox -arch:SSE2 /MT
   Platform: win32 7
       Bits: 32
       Mode: Release
************************************ System ************************************
        CPU: Intel(R) Xeon(R) CPU E3-1230 V2 @ 3.30GHz
     CPU ID: GenuineIntel Family 6 Model 58 Stepping 9
       CPUs: 8
     Memory: 15.93GiB
Free Memory: 13.03GiB
    Threads: WINDOWS_THREADS
OS Version: 6.1
Has Battery: false
On Battery: false
UTC Offset: 8
        PID: 6996
        CWD: D:\Program Files (x86)\FAHClient\work
         OS: Windows 7 Ultimate
    OS Arch: AMD64
       GPUs: 1
      GPU 0: ATI:5 Tahiti XT [Radeon HD 7970]
       CUDA: Not detected
********************************************************************************
Folding@home GPU core17

Version 0.0.55

[1] compatible platform(s):
  -- 0 --
  PROFILE = FULL_PROFILE
  VERSION = OpenCL 2.0 AMD-APP (1912.5)
  NAME = AMD Accelerated Parallel Processing
  VENDOR = Advanced Micro Devices, Inc.

(2) device(s) found on platform 0:
  -- 0 --
  DEVICE_NAME = Tahiti
  DEVICE_VENDOR = Advanced Micro Devices, Inc.
  DEVICE_VERSION = OpenCL 1.2 AMD-APP (1912.5)

  -- 1 --
  DEVICE_NAME =       Intel(R) Xeon(R) CPU E3-1230 V2 @ 3.30GHz
  DEVICE_VENDOR = GenuineIntel
  DEVICE_VERSION = OpenCL 1.2 AMD-APP (1912.5)

[ Entering Init ]
  Launch time: 2016.00.25  7:20:46
  Arguments passed: -dir 01 -suffix 01 -version 704 -lifeline 8444 -checkpoint 15 -gpu 0 -gpu-vendor ati
[ Leaving  Init ]
[ Entering Main ]
  Reading core settings...
  Total number of steps: 5000000
  XTC write frequency: 125000
[ Initializing Core Contexts ]
  Using platform OpenCL
  Looking for vendor: ati...found on platformId 0
  Deserializing System...
  Setting up Force Groups:
    Group 0: Everything Else
    Group 1: Nonbonded Direct Space
    Group 2: Nonbonded Reciprocal Space
    Found: 56372 atoms, 5 forces.
  Deserializing State...  done.
    Integrator Type: class OpenMM::LangevinIntegrator
    Constraint Tolerance: 1e-005
    Time Step in PS: 0.002
    Temperature: 300
    Friction Coeff: 0.25
  Checking core state against reference...
  Checking checkpoint state against reference...
[ Initialized Core Contexts... ]
  Using OpenCL on platformId 0 and gpu 0
  v(^_^)v  MD ready starting from step 0

Completed 0 out of 5000000 steps (0%)

Temperature control disabled. Requirements: single Nvidia GPU, tmax must be < 110 and twait >= 900

Completed 50000 out of 5000000 steps (1%)

Completed 100000 out of 5000000 steps (2%)

Completed 150000 out of 5000000 steps (3%)

Completed 200000 out of 5000000 steps (4%)

Completed 250000 out of 5000000 steps (5%)

Completed 300000 out of 5000000 steps (6%)

Completed 350000 out of 5000000 steps (7%)

Completed 400000 out of 5000000 steps (8%)

Completed 450000 out of 5000000 steps (9%)

Completed 500000 out of 5000000 steps (10%)

Completed 550000 out of 5000000 steps (11%)

Completed 600000 out of 5000000 steps (12%)

Completed 650000 out of 5000000 steps (13%)

Completed 700000 out of 5000000 steps (14%)

Completed 750000 out of 5000000 steps (15%)

Completed 800000 out of 5000000 steps (16%)

Completed 850000 out of 5000000 steps (17%)

Completed 900000 out of 5000000 steps (18%)

Completed 950000 out of 5000000 steps (19%)

Completed 1000000 out of 5000000 steps (20%)

Completed 1050000 out of 5000000 steps (21%)

Completed 1100000 out of 5000000 steps (22%)

Completed 1150000 out of 5000000 steps (23%)

Completed 1200000 out of 5000000 steps (24%)

Completed 1250000 out of 5000000 steps (25%)

Completed 1300000 out of 5000000 steps (26%)

Completed 1350000 out of 5000000 steps (27%)

Completed 1400000 out of 5000000 steps (28%)

Completed 1450000 out of 5000000 steps (29%)

Completed 1500000 out of 5000000 steps (30%)

Completed 1550000 out of 5000000 steps (31%)

WARNING:Console control signal 2 on PID 6996

Exiting, please wait. . .

wpf999

coju 发表于 2016-2-15 16:23
Project: 10468 (Run 0, Clone 494, Gen 188)

Unit: 0x0000013a538b3db9538cb7025b5b757d

这个是core17， 要core21才有

yimu35

Ars的Fahbench （GTX1080），Anand的还没出。
不知道你们测的内置监测是 Implicit or Explicit?

金鹏

yimu35 发表于 2016-5-18 10:01
Ars的Fahbench （GTX1080），Anand的还没出。
不知道你们测的内置监测是 Implicit or Explicit?
...

应该是Implicit对应PPD高低，看来内核优化到位后1080绝对百万+PPD


下图是超到1513的980TI的

yimu35

金鹏 发表于 2016-5-18 11:18
应该是Implicit对应PPD高低，看来内核优化到位后1080绝对百万+PPD

可惜zhao yutong走了后，这个bench好久没更新了吧？

金鹏

yimu35 发表于 2016-5-22 13:25
可惜zhao yutong走了后，这个bench好久没更新了吧？

有基于21内核的benchv2.2  ,置顶资料大全里有

http://fahbench.github.io/