|
|
 Author
|
Topic: optimized sources (Read 224762 times)
|
|
_heinz
|
GT540M
installed now: 304.79-notebook-win8-win7-winvista-64bit-international-beta
BOINC shows:
17.08.2012 15:05:16 NVIDIA GPU 0: GeForce GT 540M (driver version unknown, CUDA version 5000, compute capability 2.1, 1024MB, 205 GFLOPS peak)
CUDA 5.0 device driver used now.
CUDA 5.0 is available as pre release (DeveloperZone)
will run some tests in one of the next days..
_heinz
|
|
|
|
|
Logged
|
|
|
|
|
_heinz
|
I have a wattmeter and measured my machines during crunching.
The question is:How many KWh need my machines to get 1 Mio Cobblestone?
Calculated for distrrtgen
1 WU = 8758 Points
1000000 / 8758 = 114,181 ~ 115WU's needed.
Machine 1: Pentium 2.6GHz Northwood, ATI HIS Radeon HD4670 AGP
Duration per WU = 4h 10 min ~ 4,2h
115wu * 4,25h = 483h
Measured crunch-power = 140W
483h*140W=67620W ~67,62KWh
1Mio = 67,62KWh
~~~~~~~~~~~~~~~
Machine 2: Laptop i3 2.6GHz NVIDIA GT540M
Duration per wu = 2h 15min ~ 2,25h
115wu * 2,25h = 258,75h
Measured crunch-power = 62W
258,75h*62W=16042,5W ~16,0425KWh
1Mio = 16,0425KWh
~~~~~~~~~~~~~~~~~
Machine 3: ATOM 1,6GHz, ION GPU
Duration per WU = 17h 30min ~17,5h
115wu * 17,5h = 2012,5h
Measured crunch-power = 28W
2012,5h*28W=56350w ~56,35KWh
1Mio = 56,35KWh
~~~~~~~~~~~~~~~
Machine 4: V8-Xeon,2,4GHz NVIDIA GTX570, GTX470, GTX470
data from project primegrid:
per 24h = 1,6Mio
Measured crunch-power = 860W
24h*860W=20,64KWh/1,6Mio =12,9KWH per 1Mio
1Mio = 12,6KWh
~~~~~~~~~~~~~~
distrrtgen
GTX570 need ~662s/wu, 115wu*662s=76130s/3600= 21,14h/Mio
Because we have 3 GTX570 we must divide by 3
21,13h/3=7,04h per Mio
860w * 7,04h=6054,4W ~6,054KWh per Mio
1Mio = 6,054KWh
~~~~~~~~~~~~~~~
distrrtgen
GTX680 need 485,06sec/wu, 115wu*485,06s=55781,9s/3600s= 15,5h per Mio distrrtgen.
If we have 3 GTX680 we must divide by 3
15,5h/3=5,16h per Mio
We assume we need 860W with V8-Xeon.
860w * 5,16h=4437,6Wh ~4,4376KWh
1Mio = 4,4376KWh
~~~~~~~~~~~~~~~~
Remark:no CPU work done, all GPU work
I can confirm: CPU usage on all 4 machines slower than 3% (mostly 0,5-1,5)
Sure we can not direct compare primegrid against distrrtgen.
The measured results are surprizing and very interesting.
If we compare all 4 machines, V8-Xeon is the most effective.
V8-Xeon
~6,05KWh per Mio with 3 GTX570
~4,44KWh per Mio with 3 GTX680 (precalculated on real data)
As soon as I have a new PSU, I will test v8-Xeon with distrrtgen to get real data.
It is mostly effective to use modern latest hardware.
_heinz
|
|
|
|
|
Logged
|
|
|
|
|
_heinz
|
What about ATI A good sample for effectivity calculation are computers of james Ying. userid=15715lets have a look at hostid=23173Intel(R) Xeon(R) CPU E31230 @ 3.20GHz [4] AMD AMD Radeon HD 79x0 series (Tahiti) (3072MB) driver: 1.4.1720 running (distrrtgen) v3.52 (opencl_ati_101) runtime:436 sec/wu 115wu * 436 = 5140s/3600s=13,93h if we have 3GPU we divide by 3 13,93h/3=4,64h per Mio if we assume we need 860W with V8-Xeon 860W * 4,64h=3990,4 ~3,99KWh 1Mio = 3,99KWh (3 HD7970) if we have 4GPU so we divide by 4 13,93h/4=3,48h per Mio if we assume we need 860W with V8-Xeon 860W * 3,48h=2994,47 ~2,994KWh 1Mio = 2,994KWh (with 4 HD7970) This means ATI HD7970 with new OpenCL app (opencl_ati_101) can beat a GTX680 running (distrrtgen) v3.48 (cuda23) runtimerelation HD7970/GTX680:436s / 485s , difference 49sec Reference: HD7970 runtime=423s hostid=39986Intel(R) Core(TM) i5-2500K CPU @ 3.30GHz AMD AMD Radeon HD 79x0 series (Tahiti) (3072MB) driver: 1.4.1720 Reference: GTX590 runtime=848s (anonymous platform runs OCL version, maybe 2 at once?) hostid=41489Intel(R) Core(TM) i7-2600K CPU @ 3.40GHz [Family 6 Model 42 Stepping 7] [2] NVIDIA GeForce GTX 590 (1535MB) driver: 304.79 _heinz |
|
|
|
« Last Edit: 16 Sep 2012, 05:09:01 am by _heinz »
|
Logged
|
|
|
|
|
_heinz
|
Meanwhile my laptop passed 30Mio total. All GPU work. If I run cpu work additional the machine overheated. i3_gt540m_30Mio_total1Mio = 16,0425KWh 16,0425Kw/Mio * 30Mio = 481,275 KWh 30Mio = 481,275 KWh How much money cost electricity in France ? the contract has 2 prices dependant from daytime from 2:00h-8:00h and 14:30h-16:30h summary 8h(a third of a day) price= 0,0567 Euro/KWh rest of the day 16h(2 third of a day) price 0,0916 Euro/KWh so we can say 10 * 16,0425 KWh = 160,425 KWh * 0,0567 Euro/Kwh = 9,096 Euro 20 * 16,0425 KWh = 320,85 KWh * 0,0916 Euro/KWh = 29,389 Euro Summary 38,48586 Euro 30 Mio = ~38,49 Euro ~~~~~~~~~~~~~~ This is still true for the Laptop _heinz |
|
|
|
|
Logged
|
|
|
|
|
_heinz
|
100 Mio distrRTgen in sight... passed 98 Mio distrrtgenthe last two Mio will be calculated by GT540M, ATI HIS4670 AGP, ION _heinz |
|
|
|
|
Logged
|
|
|
|
|
_heinz
|
|
|
|
|
« Last Edit: 24 Sep 2012, 10:06:50 am by _heinz »
|
Logged
|
|
|
|
|
_heinz
|
|
|
|
|
« Last Edit: 17 Sep 2012, 04:09:45 am by _heinz »
|
Logged
|
|
|
|
|
_heinz
|
2nd october, shutdown all my machies, vacation  _heinz |
|
|
|
|
Logged
|
|
|
|
|
_heinz
|
vacation are over now.... 22nd november, i3 gt540m laptop get a milestone gt540m_15Mio_distrrtgen_stat and gt540m_15Mio_distrrtgenremark: its not easy to hold the GPU temp under 90 grd Celsius with a laptop, from time to time I found the machine down in the morning. gt540m_crunching_91_grd_celsius roomtemp 22 grd Celsius 22.11.2012 19:23:11 | | NVIDIA GPU 0: GeForce GT 540M (driver version 306.94, CUDA version 5.0, compute capability 2.1, 1024MB, 8381384MB available, 258 GFLOPS peak) _heinz |
|
|
|
« Last Edit: 22 Nov 2012, 03:40:44 pm by _heinz »
|
Logged
|
|
|
|
Claggy
Alpha Tester
Knight who says 'Ni!'
 Offline
Posts: 2558
|
22.11.2012 19:23:11 | | NVIDIA GPU 0: GeForce GT 540M (driver version 306.94, CUDA version 5.0, compute capability 2.1, 1024MB, 8381384MB available, 258 GFLOPS peak)
You're suffering from the Wacky Nvidia GPU Memory Bug, upgrade to Boinc 7.0.36 or 7.0.38 for the partial or full fix, But Note that since 7.0.32 and later introduce a higher internal flops value for the GPU, and which puts existing GPU tasks on the verge of going Maximum Time Exceeded, you should run down your GPU tasks prior to upgrading, Claggy |
|
|
|
|
Logged
|
|
|
|
|
_heinz
|
Hi Claggy,
I have BOINC 7.0.28x64 the actual BOINC, where I can find BOINC 7.0.38 ?
thanks in advance
_heinz
|
|
|
|
|
Logged
|
|
|
|
Urs Echternacht
Volunteer Developer
Knight who says 'Ni!'
Offline
Posts: 3066
++
|
Hi Claggy,
I have BOINC 7.0.28x64 the actual BOINC, where I can find BOINC 7.0.38 ?
thanks in advance
_heinz
http://boinc.berkeley.edu/dl/ should have all versions of BOINC. |
|
|
|
|
Logged
|
_\|/_
U r s
|
|
|
Claggy
Alpha Tester
Knight who says 'Ni!'
Offline
Posts: 2558
|
Hi Claggy,
I have BOINC 7.0.28x64 the actual BOINC, where I can find BOINC 7.0.38 ?
thanks in advance
_heinz
The Boinc 7 Changelog thread has the changes and links for the different versions after they become available (normally a day or so after they have appeared in the D/L directory, incase they implode our PC's) Claggy |
|
|
|
|
Logged
|
|
|
|
|
_heinz
|
22.11.2012 19:23:11 | | NVIDIA GPU 0: GeForce GT 540M (driver version 306.94, CUDA version 5.0, compute capability 2.1, 1024MB, 8381384MB available, 258 GFLOPS peak)
You're suffering from the Wacky Nvidia GPU Memory Bug, upgrade to Boinc 7.0.36 or 7.0.38 for the partial or full fix, But Note that since 7.0.32 and later introduce a higher internal flops value for the GPU, and which puts existing GPU tasks on the verge of going Maximum Time Exceeded, you should run down your GPU tasks prior to upgrading, Claggy I installed BOINC 7.0.38 now. 28.11.2012 09:13:48 | | Starting BOINC client version 7.0.38 for windows_x86_64 ... 28.11.2012 09:13:48 | | NVIDIA GPU 0: GeForce GT 540M (driver version 306.94, CUDA version 5.0, compute capability 2.1, 1024MB, 968MB available, 258 GFLOPS peak) 28.11.2012 09:13:48 | | OpenCL: NVIDIA GPU 0: GeForce GT 540M (driver version 306.94, device version OpenCL 1.1 CUDA, 1024MB, 968MB available) .. ION Shows now: 28.11.2012 18:34:46 | | NVIDIA GPU 0: ION (driver version 306.97, CUDA version 5.0, compute capability 1.1, 256MB, 225MB available, 53 GFLOPS peak) 28.11.2012 18:34:46 | | OpenCL: NVIDIA GPU 0: ION (driver version 306.97, device version OpenCL 1.0 CUDA, 256MB, 225MB available) Thank you both for your comment and help _heinz |
|
|
|
« Last Edit: 28 Nov 2012, 12:38:24 pm by _heinz »
|
Logged
|
|
|
|
|
|
|
|
Quote!
Any sufficiently advanced technology is indistinguishable from magic.- Arthur C. Clarke
|
 |  |  |
| |
Online users/last 15m
31 Guests, 2 Users
ML1, Pizzadude
15 Members/last 24hML1, Pizzadude, Richard Haselgrove, mr.mac52, arkayn, Urs Echternacht, Claggy, Raistmer, Mike, Josef W. Segur, Hans Dorn, Morten, Devaster, KarVi, cristipurdel
| |
 | |  |
|
