First Skylake CPUs hit the streets

Message boards : Number crunching : First Skylake CPUs hit the streets

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4

AuthorMessage
rjs5

Send message
Joined: 22 Nov 10
Posts: 273
Credit: 22,294,386
RAC: 10,897
Message 79282 - Posted: 22 Dec 2015, 16:30:31 UTC - in response to Message 79197.  

ok, this seemed to be an 'old' thread by now.
but i came across something interesting and useful

Intel power gadget
https://software.intel.com/en-us/articles/intel-power-gadget-20
the good thing is that it lets you estimate and monitor the power information in watts without needing any additional instrumentation. it depends on your CPU's model specific register. Hence, not all Intel CPUs would support that. It is apparently pretty useful if it is supported.

when my haswell i7 4771 is crunching Rosetta@hhome, crunching 8 threads concurrently

Average Processor Power(Watt)=40.1253

that would be the processor alone consumption. It won't include the energy/power consumed by RAM, harddrives etc.

i capped the cpu frequency at 3.5 Ghz, i used an aftermarket heatsink/cpu cooler http://www.coolermaster.com/cooling/cpu-air-cooler/hyper-212x/ that gives a cpu temperature of 60 deg C while the ambient temperature is 30 dec C while all 8 r@h threads are concurrently running.

when idle this processor consume a 'mere' 2-4 watts.

if you happen to run Intel power gadget for your skylake processors, do share the statistics as well



SkyLake 6700K running Fedora 23 results. I downloaded the Power Gadget for Linux and it did not support SkyLake and I would have to update it for SkyLake or wait for Intel.

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=2427886

I did not seem to make much difference over my 4GHz Haswell results. The main difference in performance beyond SandyBridge will scale with CPU Frequency and parallelism (vector). Since Rosetta is a scalar program, they will not seen measurable benefit on the newer CPU until they are able use vectors.

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=2254888



ID: 79282 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
sgaboinc

Send message
Joined: 2 Apr 14
Posts: 282
Credit: 208,966
RAC: 0
Message 79283 - Posted: 22 Dec 2015, 17:15:56 UTC - in response to Message 79282.  
Last modified: 22 Dec 2015, 17:18:58 UTC


SkyLake 6700K running Fedora 23 results. I downloaded the Power Gadget for Linux and it did not support SkyLake and I would have to update it for SkyLake or wait for Intel.

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=2427886

I did not seem to make much difference over my 4GHz Haswell results. The main difference in performance beyond SandyBridge will scale with CPU Frequency and parallelism (vector). Since Rosetta is a scalar program, they will not seen measurable benefit on the newer CPU until they are able use vectors.

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=2254888



the whestone floating point speeds for your haswell & skylake rigs seemed a little low as mine seem to get a figure closer to 4gflops
https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=1754780

i'd think it may possibly be related to frequency/turbo etc settings
or you may like to recompile the boinc client or run a recent boinc client
https://github.com/BOINC/boinc

i actually rebuild just the boinc client part with
something like
http://boinc.berkeley.edu/trac/wiki/CompileClient
./_autosetup
./configure --disable-server --enable-client CXXFLAGS="-O3 -funroll-loops -ffast-math"
make


the newer client version apparently has quite a lot of various fixes / optimization including this 'ionice' tweak for linux
https://github.com/BOINC/boinc/commit/b5a4dd390c5c2369ce7f8b668ec05c6cab5d3c45

they changed the call from

- sprintf(buf, "ionice -c 3 -n 7 -p %d", getpid());
+ sprintf(buf, "ionice -c 3 -p %d", getpid());

removing -n 7

i'm thinking that could perhaps improve the performance running memory and disk intensive jobs like r@h. -n 7 is probably the lowest priority and when combined with -c 3 (i.e. only idle), it could be too conservative to let r@h jobs run well.
ID: 79283 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
sgaboinc

Send message
Joined: 2 Apr 14
Posts: 282
Credit: 208,966
RAC: 0
Message 79284 - Posted: 22 Dec 2015, 17:52:01 UTC
Last modified: 22 Dec 2015, 17:56:12 UTC


SkyLake 6700K running Fedora 23 results. I downloaded the Power Gadget for Linux and it did not support SkyLake and I would have to update it for SkyLake or wait for Intel.

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=2427886

I did not seem to make much difference over my 4GHz Haswell results. The main difference in performance beyond SandyBridge will scale with CPU Frequency and parallelism (vector). Since Rosetta is a scalar program, they will not seen measurable benefit on the newer CPU until they are able use vectors.

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=2254888



i'm also running an active 'fancontrol' that lets me 'PWM' my cpu fan
https://wiki.archlinux.org/index.php/Fan_speed_control


it takes a bit of setup but is well worth the effort, when idle i.e. not running any jobs the fan settled at some 800rpm running quietly almost inaudible. when running boinc, it bump up to 1000-1800rpm but still pretty quiet
using an aftermarket cooler with its stock fan
http://www.coolermaster.com/cooling/cpu-air-cooler/hyper-212x/

i tend to temper down the cpu speed using 'cpupower' (e.g. cpupower -c all frequency-set -u speed ghz), to keep the temperature running around ~60deg and to run it at lower fan speed / less noise

i'd guess the 'power gadget' should work on your haswell, i'm not sure about skylake though, you may be right that the code is probably old and hence missed skylake.

do note that in linux, you need to run 'sudo modprobe msr' before 'power gadget' would work
ID: 79284 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
rjs5

Send message
Joined: 22 Nov 10
Posts: 273
Credit: 22,294,386
RAC: 10,897
Message 79285 - Posted: 22 Dec 2015, 19:37:39 UTC - in response to Message 79284.  


SkyLake 6700K running Fedora 23 results. I downloaded the Power Gadget for Linux and it did not support SkyLake and I would have to update it for SkyLake or wait for Intel.

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=2427886

I did not seem to make much difference over my 4GHz Haswell results. The main difference in performance beyond SandyBridge will scale with CPU Frequency and parallelism (vector). Since Rosetta is a scalar program, they will not seen measurable benefit on the newer CPU until they are able use vectors.

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=2254888




i'd guess the 'power gadget' should work on your haswell, i'm not sure about skylake though, you may be right that the code is probably old and hence missed skylake.

do note that in linux, you need to run 'sudo modprobe msr' before 'power gadget' would work


I will probably use liquid cooling on all systems in the future. The fan noise is effectively gone.


I used the "suggested" power_gadget options. I am not using the Intel GPU so the numbers are zero. I have a GTX 650 installed but I really hate installing the kludgy Nvidia drivers.



./power_gadget -e 1000 -d 10
System Time,RDTSC,Elapsed Time (sec),IA Frequency_0 (MHz),Processor Power_0 (Watt),Cumulative Processor Energy_0 (Joules),Cumulative Processor Energy_0 (mWh),IA Power_0 (Watt),Cumulative IA Energy_0 (Joules),Cumulative IA Energy_0(mWh),GT Power_0 (Watt),Cumulative GT Energy_0 (Joules),Cumulative GT Energy_0(mWh)
11:27:38:655,554854464118628,1.0008,3999,80.7097,80.7166,22.4213,73.8636,73.8699,20.5194,0.0000,0.0000,0.0000,
11:27:39:657,554858459077744,2.0019,3999,80.7497,161.5072,44.8631,73.9393,147.8467,41.0685,0.0000,0.0000,0.0000,
11:27:40:658,554862454347916,3.0031,3999,80.6794,242.2216,67.2838,73.8600,221.7387,61.5941,0.0000,0.0000,0.0000,
11:27:41:659,554866450025888,4.0044,3999,81.0039,323.2591,89.7942,74.1726,295.9421,82.2061,0.0000,0.0000,0.0000,
11:27:42:660,554870445209376,5.0055,3999,80.8053,404.0987,112.2496,73.9830,369.9565,102.7657,0.0000,0.0000,0.0000,
11:27:43:661,554874440960416,6.0068,3999,80.7497,484.8800,134.6889,74.0010,443.9865,123.3296,0.0000,0.0000,0.0000,
11:27:44:663,554878436639208,7.0080,3999,80.8740,565.7955,157.1654,74.0434,518.0679,143.9077,0.0000,0.0000,0.0000,
11:27:45:664,554882432438216,8.0093,3999,80.6692,646.5001,179.5834,73.8568,591.9570,164.4325,0.0000,0.0000,0.0000,
11:27:46:665,554886427758792,9.0105,3999,80.9029,727.4400,202.0667,74.0201,666.0109,185.0030,0.0000,0.0000,0.0000,
11:27:47:666,554890422490940,10.0115,3999,81.0706,808.5414,224.5948,74.2539,740.2930,205.6370,0.0000,0.0000,0.0000,

Total Elapsed Time(sec)=10.0115

Total Processor Energy_0(Joules)=808.5414
Total Processor Energy_0(mWh)=224.5948
Average Processor Power_0(Watt)=80.7611

Total IA Energy_0(Joules)=740.2930
Total IA Energy_0(mWh)=205.6370
Average IA Power_0(Watt)=73.9442

Total GT Energy_0(Joules)=0.0000
Total GT Energy_0(mWh)=0.0000
Average GT Power_0(Watt)=0.0000

TSC=554890423684984
[root@yorkfield2 power_gadget]#
ID: 79285 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
sgaboinc

Send message
Joined: 2 Apr 14
Posts: 282
Credit: 208,966
RAC: 0
Message 79287 - Posted: 23 Dec 2015, 13:45:31 UTC - in response to Message 79285.  
Last modified: 23 Dec 2015, 14:17:41 UTC


I will probably use liquid cooling on all systems in the future. The fan noise is effectively gone.


I used the "suggested" power_gadget options. I am not using the Intel GPU so the numbers are zero. I have a GTX 650 installed but I really hate installing the kludgy Nvidia drivers.

./power_gadget -e 1000 -d 10
System Time,RDTSC,Elapsed Time (sec),IA Frequency_0 (MHz),Processor Power_0 (Watt),Cumulative Processor Energy_0 (Joules),Cumulative Processor Energy_0 (mWh),IA Power_0 (Watt),Cumulative IA Energy_0 (Joules),Cumulative IA Energy_0(mWh),GT Power_0 (Watt),Cumulative GT Energy_0 (Joules),Cumulative GT Energy_0(mWh)
11:27:38:655,554854464118628,1.0008,3999,80.7097,80.7166,22.4213,73.8636,73.8699,20.5194,0.0000,0.0000,0.0000,
11:27:39:657,554858459077744,2.0019,3999,80.7497,161.5072,44.8631,73.9393,147.8467,41.0685,0.0000,0.0000,0.0000,
11:27:40:658,554862454347916,3.0031,3999,80.6794,242.2216,67.2838,73.8600,221.7387,61.5941,0.0000,0.0000,0.0000,
11:27:41:659,554866450025888,4.0044,3999,81.0039,323.2591,89.7942,74.1726,295.9421,82.2061,0.0000,0.0000,0.0000,
11:27:42:660,554870445209376,5.0055,3999,80.8053,404.0987,112.2496,73.9830,369.9565,102.7657,0.0000,0.0000,0.0000,
11:27:43:661,554874440960416,6.0068,3999,80.7497,484.8800,134.6889,74.0010,443.9865,123.3296,0.0000,0.0000,0.0000,
11:27:44:663,554878436639208,7.0080,3999,80.8740,565.7955,157.1654,74.0434,518.0679,143.9077,0.0000,0.0000,0.0000,
11:27:45:664,554882432438216,8.0093,3999,80.6692,646.5001,179.5834,73.8568,591.9570,164.4325,0.0000,0.0000,0.0000,
11:27:46:665,554886427758792,9.0105,3999,80.9029,727.4400,202.0667,74.0201,666.0109,185.0030,0.0000,0.0000,0.0000,
11:27:47:666,554890422490940,10.0115,3999,81.0706,808.5414,224.5948,74.2539,740.2930,205.6370,0.0000,0.0000,0.0000,

Total Elapsed Time(sec)=10.0115

Total Processor Energy_0(Joules)=808.5414
Total Processor Energy_0(mWh)=224.5948
Average Processor Power_0(Watt)=80.7611

Total IA Energy_0(Joules)=740.2930
Total IA Energy_0(mWh)=205.6370
Average IA Power_0(Watt)=73.9442

Total GT Energy_0(Joules)=0.0000
Total GT Energy_0(mWh)=0.0000
Average GT Power_0(Watt)=0.0000

TSC=554890423684984
[root@yorkfield2 power_gadget]#



just for your comparisons:
./power_gadget -e 1000 -d 10
System Time,RDTSC,Elapsed Time (sec),IA Frequency_0 (MHz),Processor Power_0 (Watt),Cumulative Processor Energy_0 (Joules),Cumulative Processor Energy_0 (mWh),IA Power_0 (Watt),Cumulative IA Energy_0 (Joules),Cumulative IA Energy_0(mWh),GT Power_0 (Watt),Cumulative GT Energy_0 (Joules),Cumulative GT Energy_0(mWh)
21:34:26:319,8088984507242,1.0007,3700,57.7180,57.7228,16.0341,48.0270,48.0309,13.3419,0.6968,0.6968,0.1936,
21:34:27:320,8092479094988,2.0015,3700,56.0639,113.8039,31.6122,47.0419,95.0873,26.4131,0.0362,0.7331,0.2036,
21:34:28:321,8095973487255,3.0022,3699,56.8743,170.6882,47.4134,47.0844,142.1799,39.4944,0.8167,1.5499,0.4305,
21:34:29:322,8099468026311,4.0029,3700,56.6179,227.3179,63.1439,47.5436,189.7334,52.7037,0.0511,1.6010,0.4447,
21:34:30:322,8102962605109,5.0036,3700,57.6191,284.9494,79.1526,48.5990,238.3429,66.2064,0.0328,1.6339,0.4538,
21:34:31:323,8106456984079,6.0043,3700,57.9990,342.9611,95.2670,48.8540,287.2076,79.7799,0.0349,1.6688,0.4635,
21:34:32:324,8109951193295,7.0049,3700,58.1156,401.0866,111.4129,49.0365,336.2524,93.4035,0.0311,1.6999,0.4722,
21:34:33:324,8113445474575,8.0056,3700,59.5958,460.6932,127.9703,50.5130,386.7745,107.4374,0.0339,1.7338,0.4816,
21:34:34:325,8116939702223,9.0062,3700,59.7040,520.4077,144.5577,50.6236,437.4070,121.5019,0.0334,1.7672,0.4909,
21:34:35:326,8120433920500,10.0069,3699,59.6727,580.0910,161.1364,50.6019,488.0179,135.5605,0.0348,1.8020,0.5006,

Total Elapsed Time(sec)=10.0069

Total Processor Energy_0(Joules)=580.0910
Total Processor Energy_0(mWh)=161.1364
Average Processor Power_0(Watt)=57.9692

Total IA Energy_0(Joules)=488.0179
Total IA Energy_0(mWh)=135.5605
Average IA Power_0(Watt)=48.7683

Total GT Energy_0(Joules)=1.8020
Total GT Energy_0(mWh)=0.5006
Average GT Power_0(Watt)=0.1801

this is without frequency throttling and running 8 concurrent r@h threads
temperature is hovering around 60-65 deg C, ambient 30 deg C

apparently Total IA power seemed to be the power used in the cores i guess and GT power seemed to be the graphics core. I'm using the on-chip GPU (no external cards)


i'd think your CPU is actually quite a bit more performant running r@h hence the lower Whestone gflops is a little surprising.

i'm not sure if it may be the cpu governer that you are running, i.e. it cranks up the frequency a little too late while the Whestone benchmark is running.
i'd guess you could try running the benchmark twice / thrice in quick succession in boinc client and it may show the actual Whestone gflops that your CPU is running at :)

i think the Whestone gflops is actually a poor indicator of the real performance of rosetta@home, unfortunately boinc is really measuring Whestone credits/gflops rather than rosetta@home credits/gflops,
more than likely there are code optimizations by the compiler in r@h code that's not there in the whestone benchmark and vice versa :o :p lol
https://boinc.bakerlab.org/rosetta/forum_thread.php?id=6744
ID: 79287 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Previous · 1 · 2 · 3 · 4

Message boards : Number crunching : First Skylake CPUs hit the streets



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