Duvie's Dual Dual-core 270 benches

[Duvie]

System specs:

(2) Opteron 270 (dual core) = 4x cores
Iwill DK8X Amd 8131 motherboard
2gb (4x512) Corsair XMS PC3200LL (cas 2-3-2-6) Dual channel per cpu
(2) WD 250GB SATA II HDD in RAID 0 (64 stripe)
Gainward 2400Ultra (6800GT) 412/1120 converted to Quadro FX4000
600watt OCZ Powerstream
WinXP Pro w/ SP2 & AMD Cpu driver


Think of

1 core = >3200+ A64 Rev E 90nm
2 core = >3800+ X2

should be comparable. L2 cache has limited effects in the apps I will be testing.

RAID 0 performance

http://home.comcast.net/~duvie23/DUAL270-atto-250raid0.JPG
http://home.comcast.net/~duvie23/DUAL270-atto-250raid0-8mb.JPG

TMPEG 2.52

First off I decided to convert a Hidef AVI file to MPEG2 DVD standard....Clip is a Trailer from Xmen 3...

First test was to use Automatic Variable Bitrate (setting 80 hi6000/lo2000) 8 bit and normal motion precision.

I used affinity to test 1,2,3,&4 cores


1 core = 3:45
2 core = 2:00
3 core = 1:44
4 core = 1:44 (2.16x faster 1 core, 1.15x faster then 2 cores)

http://home.comcast.net/~duvie23/tmpeg-test-AQ-2cores.JPG
http://home.comcast.net/~duvie23/tmpeg-test-AQ-3cores.JPG
http://home.comcast.net/~duvie23/tmpeg-test-AQ-2cores.JPG

As you can see dual cores would have been cpu limited in this test. Watching cpu graphs in task manager I routinely saw usage spike to 64-70....Sincei t did not spike over 75 we see there is no benefit for the 4th core....It might as well be Folding or doing some background stuff....



Second test was to use 2 pass Variable Bitrate (setting 80 hi6000/lo2000) 8 bit and normal motion compensation.

I used affinity to test 1,2,3,&4 cores

1 core = 7:12
2 core = 3:56
3 core = 3:27
4 core = 3:28 (2.07x faster 1 core, 1.13x faster 2 cores)

As you can see similar pattern. No extra need or use of the 4th score. I witnessed similar cpu usage as above. As you will see in the next test this is likely do to quality setup of motion search precision.



Third test was to use Automatic Variable Bitrate (setting 80 hi6000/lo2000) 10 bit component precision and High Quality motion precision. Merely higher quality finished product.


I used affinity to test 1,2,3,&4 cores

1 core = 5:43
2 core = 3:00
3 core = 2:03
4 core = 1:53 (3.03x faster 1 core, 1.59x faster 2 cores)

http://home.comcast.net/~duvie23/tmpeg-test-AQ-highqual-2cores.JPG
http://home.comcast.net/~duvie23/tmpeg-test-AQ-highqual-3cores.JPG
http://home.comcast.net/~duvie23/tmpeg-test-AQ-highqual-4cores.JPG

As you can see there was a gain by using 4 cores as often cpu usage was spiking in the 88-80% range. The added high quality with motion and DC precision facilitated the extra need for cpu processing power.



I tried a 4th test for Highest quality of motion precision and got similar graph numbers as just High Quality....So I didn't progess on after testing 3 cores...

I also tried a 5th test to see if theory holds true the other way. Chooing 8 bit DC precision and Lowest quality (very fast) motion precision. What I saw was a cpu usage that never cracked 50% and therefore would have shown no gain after 2 cores....Actual average cpu usage was like 40% so therefore the added speed increase over a single core was ni the 80% range.

2 INSTANCE of TMPEG 2.52

I used Automatic Variable Bitrate (setting 80 hi6000/lo2000) 8 bit and normal motion precision.

I created another directory and labelled the executable TMpegenc-A. I copied the avi file into another directory to have 2 separate and unique names to read from.

1 instance 2 core = 2:00
1 instance 4 core = 1:44
2 instances 4 core = cores 1-2 = 1:53 / cores 3-4 = 1:43

http://home.comcast.net/~duvie23/2instance-tmpeg-during.JPG
http://home.comcast.net/~duvie23/2instance-tmpeg-final.JPG


Cpu usage was not pegged at 100% with some spiking in cpu 1-2 and thus the slower speed...I am not sure if I was IO limited in that bit, but since 2 cores did that same file with same settings at 2:00 it appeared as I actually got more done...

People ponder that one a bit...I am not sure what to make of it yet...

TMPGEnc 3.0 Xpress

Using the sameclip as above a (~2min 30sec clip)

1st test:

I used 10bit DC, High for motion search precision. Same VBR bitrate as above and 2pass...no autmotic VBR in this version...

prioritize speed (was the only way I could get the usage above a 55-72% range)

1 core = 9:26 (pegged at 25%)
2 cores = 5:03 (45-50% range)
4 cores = 2:49 (80-95% range) (3.34x faster 1 core, 1.79x faster then 2 core)

Now for something strange....Dropping the quality settings to 9bit DC and standard motion precision here are the times...

1 core = 6:32
2 cores = 3:40
4 cores = 2:43

1 core dropped it times tremendously as it obviously was not doing the quality work...same for 2 cores though not as much and the pcu usage was hovering in the 44-50% range...not pegged....

4 cores was only slightly difference but the program would never tax the 4th cpu....It ran in the 55-72% range....It is like the damn thing paces itself....

The only way to get full speed is to max quality, max filters (which usually are not needed with a quality source), and set prioritize speed....

To show this...

2nd test:

I used same settings of DC, motion, and bitrate...Only difference is Priority quality.

2 cores = 4:34 (47-50% usage)
4 cores = 3:43 (55-72% usage) never uses more then a fraction of the 3rd core)

only 22% faster....reviewing the video and no noticeable quality difference is present...



CINEBENCH 9.5(32 bit version)

I used affinity to test 1,2,3,&4 cores

1 core = 295
2 core = 543
3 core = 772
4 core = 979 (3.32x 1core, 1.80x 2core, 1.27x 3core)

http://home.comcast.net/~duvie23/DUAL270-cinebench-1core.JPG
http://home.comcast.net/~duvie23/DUAL270-cinebench-4cores.JPG

As you can see Cinebench scales great with multiple cores...If you are not familar with how it looks when it renders it with multiple cores, look here...

http://home.comcast.net/~duvie23/DUAL270-cinebench-render.JPG

In this example the top 1/4 gets completed far faster so then it then jumps down and splits the avaiable opening between the 3rd and 4th quarter and conitues working....



GORDIAN KNOT w/ H.264 CODEC

I used a 60sec trailer from the DVD "gone in 60 seconds" (like a 75-80mb file) compressed it down to 18mb, auto crop, 2 pass multipass with default quality and target nbitrate a default of 800 in 1st test and 2000 in 2nd test...

I used affinity to set cores but also set the number of threads in the H.264 codec setup

1st test

1 cores (1 thread) = 5:36
2 cores (2 threads) = 3:22
3 cores (3 threads) = 2:44
4 cores (4 threads) = 2:13 (2.52x faster 1 core, 1.52x faster 2 cores)

2nd test

1 cores (1 thread) = 6:39
2 cores (2 threads) = 3:52
3 cores (3 threads) = NA
4 cores (4 threads) = 2:19 (2.87x faster 1 core, 1.66x faster 2 cores)

I have all the logs I iwll list them....

It appears as quality mode increases the extra cores will be more taxed....Odd thing was that excpet for 1 core 1 thread the other cores were never maxed...1 core stayed 25%, but 2 cores was like avg 44-45%...3cores was like 62-75%...

http://home.comcast.net/~duvie23/DUAL270-Gknot-H.264-test1-4cores-during.JPG

Here is a pic during test1 with 4 cores/4 threads...I saw basically from mid 60's to low 80's for usage....with test 2 i saw peaks in the high 80's and less 60's...It obviously used a bit more...

****ADDED 3RD TEST*****

I used higher quality settings thanks to a more knowledgeable H.264 member who knew the settings to make it....bitrate is same as test1...still used affinity to set cores used as well as # of threads in the codec...

1 core = 10:42 (first pass alone was 3:59)
2 cores = 5:52
4 cores = 3:39 (2.93x faster 1 core, 1.61x faster 2 cores)





DVDSHRINK 3.2

I used Gone in 60 Seconds DVD. I ripped to HDD with DVD-decryptor to isolate cpus. I did a full back up less foreign languages...It is a 65% compression to keep it to DVD 4.7gb standard. I use sharp adaptive. I used Deep anaylsis the first time, but after you run it once it wont do it again...It caches the info someplace I couldn't find....


4 cores w/o deep analysis = 15:40 (3.07x faster 1 core, 1.59x faster 2 cores)
4 cores w/ deep analysis = 21:01
4 cores w/ deep analysis (ripped from DVD) = 41:29 (analysis took 20min by itself)
4 cores w/o deep analysis (ripped from DVD) = 21:46

3 cores w/o deep analysis = 16:38 (1.50x faster 2 cores)

2 cores w/o deep analysis = 24:54

1 core w/o deep analysis = 48:10

As you can see it can use the cores and is very well multithreaded. the problem may be for many the IO limitation of the DVD-rom and the ripping speed. Not all DVD-roms are the same and some have ripping locks on them. This will definitely have the cpu in wait mode. Remember I am also using a RAID setup...

As you can see in these first screenshots 4 cores in analysis and encoding definitely uses more then 3 cores...

http://home.comcast.net/~duvie23/DUAL27...dshrink-readingfilesfromHDD-4cores.JPG
http://home.comcast.net/~duvie23/Dual270-dvdshrink-4cores-deepanalysis-usage.JPG
http://home.comcast.net/~duvie23/Dual270-dvdshrink-4cores-encoding-usage.JPG
http://home.comcast.net/~duvie23/Dual270-dvdshrink-4cores-finalwoanalysis.JPG

in the next screen shot you see 2 cores is fully pegged and is obviously cpu limited...

http://home.comcast.net/~duvie23/Dual270-dvdshrink-2cores-woanalysis-usage.JPG


ADT2004 ARCHITECTURAL DESKTOP 2004

Did some animation test here of several walk-thrus, flyarounds and stills...Some will play with highere resolutions, higher quality, and radiosity....results are also interesting...

1st test:

Simple walk-thru (0 to 33 frames) at 320x240 converted to 1 pass Xvid (high quality)

1 core = 7:03
2 cores = 4:17
4 cores = 3:56

Now I used H.264 codec and set appropraite thread amounts

1 core = 7:02
4 cores = 3:55

As you can see not much different. Minimal less then 10% gain for 4 cores over 2 cores...not very taxing. 2 cores showed a heality gain though of 65% over 1 core.

Now I used Xvid again but made final size 640x480...

2 cores = 14:44
4 cores = 13:31

Again only 9% gain....Resolution made no differnce here Maybe 640x480 is still not taxing enough

Next I used 1024x768 with Xvid (to speed up time I used the first 3 frames of 0 to 33 above) with same XVid settings.

1 core = 5:10
2 cores = 2:55
4 cores = 2:28

Here we see the 4 cores building some separation with an 18% faster time then 2 cores, and 4 cores is 109% faster then 1 core. Still not the best but it seems that has resolution become larger it uses more of the multi-processors...



2nd Test:

I used a High res 1900-1200 to render a detailed interior of a Library complete with lights and shadows. (radiosity already calculated so it doesn't have to do it every time you rerender it)

1 core = 2:23
2 cores = 1:20
4 cores = 1:10

A little bit better here as 4 cores was 14% faster then 2 cores and finally doubled the speed of 1 core.


3rd test:

Next I used frames 100-102 which were of a study/den with detailed wood floors, high reflect black surface of a piano ect. I used radiosity....rendered at 1024x768

Radiosity time alone.

2 cores = 5:11
4 cores = 2:36

It appears that computing radiosity geometry was 2x time faster then 2 cores and pegged usage at 100% as will be seen in another example below...

Rendering the frames after that showed 30% gain over 2 cores and 140% over 1 core

The more frames you have to compute radiosit the 4 cores will definitely speed this up...


4th test:

I used an exterior image with ran radiosity to calculate materials, and sun lighting and shadows. This one is not as highly detailed since there are less reflective materials visiable. I rendered the image to a res of 1200x900 w/ 60% radiosity.

radiosity alone

1 core = 4:05
2 cores = 2:54
4 cores = 2:01

As you can see it was 43% faster then 2 cores and 100% faster then 4 cores...

Rendered time for frame

1 core = 5:49
2 cores = 3:55
3 cores = 2:55

34% faster then 2 cores and rihgt at 100% faster then 4 cores...


Conclusion ADT 2004 is not as multithreaded as I hoped except when like encoding you start piling on the higher quality, higher res, and radiosity.

I do need to do an image with multiple lighting sources..

http://home.comcast.net/~duvie23/DUAL270-ADT-Hq1024-1core.JPG
http://home.comcast.net/~duvie23/DUAL270-ADT-Hq1024-2core.JPG
http://home.comcast.net/~duvie23/DUAL270-ADT-Hq1024-4core.JPG

http://home.comcast.net/~duvie23/DUAL270-ADT-Hq1200-1core-usage.JPG

http://home.comcast.net/~duvie23/DUAL270-ADT-Hq1200-2core-rad.JPG
http://home.comcast.net/~duvie23/DUAL270-ADT-Hq1200-2core-usage.JPG

http://home.comcast.net/~duvie23/DUAL270-ADT-Hq1200-4core-Rad.JPG
http://home.comcast.net/~duvie23/DUAL270-ADT-Hq1200-4core-usage.JPG

 

[RaynorWolfcastle]

Interesting... I have a request, could you compare 2 cores against 4 coes in H.264 compression? I's be interested in seeing how x264 encoding performance scales.

 

[DrMrLordX]

Interesting. This should give us an idea of how much further the multicore approach can go with existing applications. Once you hit 4 cores, you're probably not going to get any better performance from a single multithreaded app(at least not from TMPEG 2.2 anyway).

 

[Duvie]

Originally posted by: DrMrLordX
Interesting. This should give us an idea of how much further the multicore approach can go with existing applications. Once you hit 4 cores, you're probably not going to get any better performance from a single multithreaded app(at least not from TMPEG 2.2 anyway).

I think only higher quality processing will require mor cpu power so perhaps after I test H.264 stuff we will see what the future holds for multicore in high end video processing...

If anything 2 average setup encoding could be done simultaneously...each using 2 cores and having all the power of 2 machines....I may bench that theory...good idea!!!


edit: I added a test with 2 instances of TMPGenc to show if the program can not make effective use of all cores you could do twice the work with surprising results...

 

[bjc112]

Keep them coming D.

 

[Duvie]

Working on a H.264 test...May take me awhile...I am not use to this new and fast moving format...

 

[stevty2889]

Very nice. Looks like there will be plenty of uses for a quad core.

 

[Duvie]

http://home.comcast.net/~duvie23/DUAL270-4CORES.JPG

Ofcourse this is what it will do when I am not doing more pressing things...4boxes in one

 

[AkumaX]

awww... too bad TMPGenc only scales with higher quality modes i would have loved to see a direct performance increase when going from 1-2-4 cores.
i see you're using an older version of TMPGenc, maybe they increased multi-threading performance in TMPGenc3Express?

 

[Duvie]

Originally posted by: AkumaX
awww... too bad TMPGenc only scales with higher quality modes i would have loved to see a direct performance increase when going from 1-2-4 cores.
i see you're using an older version of TMPGenc, maybe they increased multi-threading performance in TMPGenc3Express?

javascript:winopen('messagepost.aspx?postaction=reply&catid=28&threadid=1830732&messid=22835882&STARTPAGE=1&ezquote=y&parentid=22835882','msgrep22835882',710,550);
Quote

I dont have a copy of that...This one I own....If there is a trial version send me to it.....

 

[Duvie]

H.264 has started testing....with average setup I hit over 75% quite a bit with an aveareg in the low 70%...So it looks like at normal quality it is a 3 core machine by 4cores can and should show 50% over dual core alone....

I will put up charts....

I will turn up the quality as well as go for more of a compression...

 

[PetNorth]

nice system.

Here you have http://www.pegasys-inc.com/en/download/te3xp.html#trial

can you test Premiere Pro 2.0 and Vegas 6?
Lightwave 8? 3dsmax 8?

 

[Duvie]

Originally posted by: PetNorth
nice system.

Here you have http://www.pegasys-inc.com/en/download/te3xp.html#trial

can you test Premiere Pro 2.0 and Vegas 6?
Lightwave 8? 3dsmax 8?

i will do the best I can....I have a full licensed copy of 3dsmax6 and downloaded the trial of 7 about 12 months ago....If they have a trial there I will do it...I am sure I have premiere pro....No lightwave 8 or vegas 6 though...

 

[PetNorth]

yes, there is a 3dsmax 8 trial, but version 6 or 7 are fine anyway
Vegas 6 has a trial too http://www.sonymediasoftware.com/download/step2.asp?DID=583

 

[Duvie]

Originally posted by: PetNorth
yes, there is a 3dsmax 8 trial, but version 6 or 7 are fine anyway
Vegas 6 has a trial too http://www.sonymediasoftware.com/download/step2.asp?DID=583

Cool...

I put up H.264 codec test using Gknot...probably not the most taxing sense it was a trailer....I didn't have the time ot do a full movie...I may take a VOB from Star Wars and clip a piece with DVD2AVI and do it that way...I am sure the source is far better and higher initial bitrate...

 

[bunnyfubbles]

Ok, I'm only a little jealous

 

[Megatomic]

Hey, I posted in this thread last night? What happened to my post? I even added it to my favorites and that link disappeared.

Anyways, here are my 4800+ @ 2.5GHz Cinebench 9.5 results:

Cinebench_Result.JPG

I thought my machine would be more competitive.

 

[Duvie]

That is right in line with my 706 I got with my 4400+ at 2.6ghz....

 

[Duvie]

DVDshrink benches added!!!

 

[Hard Ball]

Originally posted by: Duvie
DVDshrink benches added!!!

Hey Duvie,

Would you be willing to test a small bit of code on your machine, of which I'm suspicious that the run time is not right. I'm just testing to see how well (or if) python interpreter can do with multiple threads. And since we have very similar machines, if you have linux installed, you can easily test to see whether it's an issue with python utilizing hardware SMP in general, or simply the specific setup on my machine.

let me know

Thanks
HB

 

[Avalon]

Good results Duvie!

 

[rogue1979]

This review is clearly second fiddle......

No benchmarks to tickle my fancy!!!!!!





Just kidding you Duvie, don't get upset

 

[Duvie]

Originally posted by: Hard Ball

Originally posted by: Duvie
DVDshrink benches added!!!

Hey Duvie,

Would you be willing to test a small bit of code on your machine, of which I'm suspicious that the run time is not right. I'm just testing to see how well (or if) python interpreter can do with multiple threads. And since we have very similar machines, if you have linux installed, you can easily test to see whether it's an issue with python utilizing hardware SMP in general, or simply the specific setup on my machine.

let me know

Thanks
HB

I do not have linux installed and know nothing about the software...i was thinking about dual booting a 64bit XP but that is it... so sorry!!

I will test compiling for ppl all you have to do is send me the app the code and instructions what to do...

 

[Hard Ball]

Originally posted by: Duvie

Originally posted by: Hard Ball

Originally posted by: Duvie
DVDshrink benches added!!!

Hey Duvie,

Would you be willing to test a small bit of code on your machine, of which I'm suspicious that the run time is not right. I'm just testing to see how well (or if) python interpreter can do with multiple threads. And since we have very similar machines, if you have linux installed, you can easily test to see whether it's an issue with python utilizing hardware SMP in general, or simply the specific setup on my machine.

let me know

Thanks
HB

I do not have linux installed and know nothing about the software...i was thinking about dual booting a 64bit XP but that is it... so sorry!!

I will test compiling for ppl all you have to do is send me the app the code and instructions what to do...

Thanks a lot,

Here is the code (although the indentation is not right, which must be aligned for python scripts);

import sys, re, time
from threading import Thread as T

class ST(T):

def __init__(self, start, loop):
self.start = start
self.loop = loop
self.list = []

def run(self):
res = fact(self.start, self.loop)
self.list.append(res)

class MT(T):

def __init__(self, new_obj):
self.obj = new_obj

def run(self):
pass

def fact(base, x):
res = base
for i in range(0, x, 1):
base += 1
res *= base
print res
return res


for x in range(0, 10, 1):
newlist = []
new1 = ST(1, 10000)
newlist.append(new1)
new2 = ST(1, 10000)
newlist.append(new2)
new3 = ST(1, 10000)
newlist.append(new3)
new4 = ST(1, 10000)
newlist.append(new4)
for new in newlist:
new.run()

I can get the code to you another way; do you have cygwin so you can ssh into my workstation? or I could just email you the code. Let me know

Thanks
--HB

 

[JackBurton]

You're the man Duvie! Nice machine and great benchmarks. I'm trying to hold off my upgrade until quad core CPUs come out.