PDA

View Full Version : PRMAN performance on a INTEL and AMD box


orbitalpunk
11-06-2004, 04:47 AM
I have seen tons of benchmarks comparing mental ray and maya renderers on xeons and opterons but none could be found with pixar's renderman. would someone with dual Xeons benchmark with me. im still looking for a scene to use but if someone has one id be willing to use it.

thanks

BillSpradlin
11-06-2004, 08:29 AM
Mental ray loves AMD Opterons, but the standard AMD Chips don't perform as well as the Intel P4's and Xeons. However a dual Opteron 250 box will smoke the living bejesus out of a Dual 3.4ghz 64bit Xeon machine. That's pretty nifty considering the Opterons are 1.2ghz slower in CPU speed.

As far as Renderman goes, I really couldn't tell you off the top of my head as I don't have an AMD machine to test it on.

orbitalpunk
11-06-2004, 11:46 PM
Hi there,

Thanks for postin.

well it seems opterons take a beating at zoorenders mental ray benchmark page

http://www.zoorender.com/html/benchmark_mental.htm

the opterons seem accurate cause im clocking in at 47 seconds myself on my dual opteron 248.

and thats what has made me think if PRMAM would perfrom as well also due to multi threaded abilites thru HyperThreading. Do you have PRMAN? would be able and whiling to do a test with me? we can find a mutual scene and see what we come up with. what are you machine specs? Im right now running on a dual opteron 248. but considering changing over to a dual 3.4 xeon depending on my findings with PRMAN.

gga
11-07-2004, 10:48 AM
Short answer is probably no. Prman so far is a single threaded application (note that this is not the case of other renderman renderers like 3ddelight or air). As such, you should see pretty much no improvement due to hyperthreading.
Note, however, that latest versions of prman can use up to two cpus thru a licensing scheme similar to netrender, which will be useful for dual cpu machines.

In terms of Intel vs AMD... you will have to test it for yourself as it varies with time.
Pixar's biggest customers are most likely AMD based, but Intel engineers were subcontracted by Pixar to do microoptimizations of the software a year or so ago.
If you are to believe all press releases, Pixar's (Animation) renderfarm was, until 2003, heavily intel based with AMD cpus used mostly for servers. Recent reports from Pixar indicate Apple machines becoming their preferred desktop boxes.

orbitalpunk
11-08-2004, 12:14 AM
thank your input gga. its the best ive hear so far. seems like pixar rendman folks are dormant on the subject. I email pixar and am waiting for a reply. i also got a reply from zoorender and were gonna work out a test and he'll put it up his site. i just hope i can get some more answers soon cause i have about 2weeks left to return my stuff if i wanna swtich :P

beaker
11-08-2004, 05:46 AM
I'm not sure where your buying machines from but Boxx for example you could probably get a demo p4 and opteron machine along with demo licenses of prman from Pixar so you can do the tests yourself.

orbitalpunk
11-08-2004, 07:34 AM
true, but thats a lot of work. physical and financial.

Ed Caracappa
11-08-2004, 01:54 PM
If someone has a scene, set up to render on renderman (all files including shaders and textures, etc) and wants to send it to me I'll run a quick test.


Ed

dmaas
11-08-2004, 03:51 PM
Ed I might just take you up on that. What is your email address?

I am just about to take delivery of a dual 3.2GHz Xeon machine. I would be happy to run some PRMan benchmarks since I am curious about the performance relative to AMD machines too. I also have access to single-CPU 3GHz P4s but no Opterons.

(where are people buying dual Opterons these days? I am having a hard time finding them for prices comparable to dual Xeons... Got mine from Dell for ~$4000)

Also note that PRMan's pricing distorts the performance picture... It's probably more cost-effective to buy expensive, fast dual-CPU machines than lots of slower single-CPU machines when you factor in the license costs. This means that Mac G5s might actually be more cost-effective than AMD or Intel machines. (haven't done any benchmarks though)

Ed Caracappa
11-08-2004, 03:53 PM
ed@boxxtech.com

512-225-6313 direct line

dmaas
11-08-2004, 11:19 PM
I made up a little RenderMan benchmark. See here:
http://www.dcine.com/prman-bench/

orbitalpunk
11-09-2004, 01:54 AM
Hey dmass, thanks for trying to put up a scene. unfortuantely i have no idea how to port that to my windows machine. if anyone is successfull ill run it asap. i can do tests on a dual 248 opteron machine. i might post up a scene today tho that comes from the RAT tutorials. just wanna complicate it a bit more so i takes more time to render. right now it finishes in 7 seconds.. ill try to kick up to 15+ with some refractions or somthing. I was gonna use the caustics tutorial but im having problems getting the caustics to work. Ive gone over my shader settings so many times. argh..

i also did get a reply from pixar. and here is what they had to say -

We have a number of customers running both Intel and AMD
boxes, and both seem happy with results. I will point out
a couple of things, however:

* The PRMan applications do indeed use SSE2 code, and
it is entirely possible it will use SSE3 code in the future.

* All of our development for our 32 bit releases (windows,
and linux), are all done currently on Intel hardware. Our
64 bit release is compiled on AMD hardware.

We do not recommend hyperthreading, as we've found that it
actually degrades rendering performance.

dmaas
11-09-2004, 02:19 AM
I just added a .ZIP format archive which you can use on Windows.

The "dobench" script won't work, but you can just compile the shader and render the RIBs manually. i.e. make a DOS prompt at the directory where the files are, then run "shader bench.sl" to compile the shader and "prman -progress -statistics 1 raydiff.rib" to render and time a RIB. (this assumes the RenderMan executables are in your PATH environment variable; if not, you'll have to specify the path like "C:\Program Files\RenderMan Pro Server\bin\shader.exe" bench.sl etc...

(don't bother rendering geom.rib, it's just an archive used by the other RIBs)

Here are my results on two systems:OS: Linux 2.4.28-pre3
# of CPUs: 2 (including HyperThreading virtual CPUs)
CPU type: Intel(R) Pentium(R) 4 CPU 3.06GHz
CPU speed: 3049.540 MHz
CPU cache: 512 KB

Shader time: 388.69 sec
Hider time: 576.24 sec
Diffuse raytrace time: 421.98 sec
Specular raytrace time: 390.77 sec

OS: Linux 2.4.28-pre3
# of CPUs: 2 (including HyperThreading virtual CPUs)
CPU type: Intel(R) Pentium(R) 4 CPU 3.00GHz
CPU speed: 2992.553 MHz
CPU cache: 1024 KB

Shader time: 439.91 sec
Hider time: 597.98 sec
Diffuse raytrace time: 479.44 sec
Specular raytrace time: 456.29 sec

BillSpradlin
11-09-2004, 03:16 AM
"Recent reports from Pixar indicate Apple machines becoming their preferred desktop boxes."

Heh, don't believe everything you read. Sometimes press releases are only there to please the CEO/marketing giants, and bear little to no factual truth behind them.

orbitalpunk
11-09-2004, 04:35 AM
my prman doesn't have the statistics option for some reason. did they change its nameing by any chance. these are the options available when using help

debug, revcover 0:1, progress, version, catrib ribfile, rif plug-in

well im running it now. is there any other way to get the stats?

orbitalpunk
11-09-2004, 04:47 AM
hmm, i ran it and only took like 3 minutes. did i miss something? here is the tiff file render (converted to jpg for this display)

http://members.dslextreme.com/users/orbitalpunk/raydiff.jpg

http://members.dslextreme.com/users/orbitalpunk/radiff.jpg

dmaas
11-09-2004, 09:43 AM
Oops, my bad. I think I was imagining the -statistics switch.

Just add this at the top of the RIBs:Option "statistics" "endofframe" [1]

The number you want is "real time" at the top.

orbitalpunk: that's correct. I wanted the test scenes to render in less than 10 minutes, so the raytrace scenes are a little rough. Can you let me know your processor type/speed and exactly how long it took to render in seconds?

orbitalpunk
11-10-2004, 02:04 AM
oook, just finished for the second time testing. man, i was shitting it when i say my first results. they were very disapointing. but then i noticed thru taskmanager i was only using 1 cpu. seems my renderman.ini didnt have the mulitprocessor line like its supposed by default. ill post here tests with single and dual processors.

by the way, why are some of your times so different between the two. cpus are only 50mhz appart, right? and im surrprised your 3.0 lost to the 3.06. isnt that 3.0 the nacona core with 800FSB? still, i didnt think 50+ mhz would make such a difference.

say, anyone got a dual 3.4 that wants to benchmark with me? price wise, that would be my competition.

system specs:
Dual AMD Opteron 2.2ghz
1GB RAM DDR400 2.5CAS
WindowsXP SP2 (non-NUMA)

Single CPU:

Shader 455 seconds
http://www.orbitalpunk.com/rendermangifs/shader.gif

Hider 591 seconds
http://www.orbitalpunk.com/rendermangifs/hider.gif

Raydiff 331 seconds
http://www.orbitalpunk.com/rendermangifs/raydiff.gif

Rayspec 320 seconds
http://www.orbitalpunk.com/rendermangifs/rayspec.gif


Dual CPU:

Shader 242 seconds
http://www.orbitalpunk.com/rendermangifs/shader2.gif

Hider 335 seconds
http://www.orbitalpunk.com/rendermangifs/hider2.gif

Diffuse Raytrace 246 seconds
http://www.orbitalpunk.com/rendermangifs/raydiff2.gif

Specular Raytrace 219 seconds
http://www.orbitalpunk.com/rendermangifs/rayspec2.gif

dmaas
11-10-2004, 02:30 AM
Thanks for the results! I will post these up on my website - I think it will help people make buying decisions.

The difference between my two 3GHz Pentium 4s is the RAM. The faster one uses RAMBUS RAM (!), the slower one uses some kind of DDR SDRAM (533MHz FSB I think - not the very latest). I too was surprised at the difference, but it's true - the RAMBUS system outperforms the DDR system on my production shots by about 10-20%.

I will post scores for a dual 3.2GHz Xeon (800MHz FSB) as soon as it arrives.

I would LOVE to see someone post scores from a Mac G5...

orbitalpunk
11-10-2004, 03:21 AM
ill ask barefeats.com . i was also trying to get a 64bit under windows 64 but pixar only designed it under there linux version. you think yoiu could try it with your version of prman? its in 11.5.3

P.S. Firefox 1.0 is out! so is konfabulator for windows!
http://www.konfabulator.com/
http://mozilla.org

dmaas
11-10-2004, 04:34 AM
I've put up a "scoreboard." Anyone else is welcome to contribute, just send me your times.

I will try running PRMan in both 32-bit and 64-bit modes on my new system.

I just checked on my DDR system - the invoice says it has DDR400 RAM. No idea why it is so slow. Maybe it's the Intel motherboard (an "entry-level" server board).

orbitalpunk
11-10-2004, 05:04 AM
are you sure both cpu's are running. either have the /prman/nprocessors 0 line in the renderman.ini file or use the -p switch. just some info for anyone else as well.

orbitalpunk
11-10-2004, 05:08 AM
barefeats is interested in benchmarking with us but he doesnt own renderman. but i told him they do have a demo of it.. but im not sure where to get it. but i have seen it in there FAQ. does anyone have more info on how to get the demo? do i need to email renderwoman?

thanks

dmaas
11-10-2004, 07:31 AM
Yes you have to email Pixar to get a demo license... They want to get to know you before letting you use their software :)

BTW the Pentium scores on my site are all single-CPU. I don't have the dual system yet.

dmaas
11-16-2004, 12:24 AM
For those of you following this, the Opteron 250 is pretty much kicking everyone's ass.

The 3.2GHz Xeon is a little slower than the Opteron 248. The 3.6GHz Xeon might beat an Opteron 248 but would not likely outperform an Opteron 250.

The 2.5GHz Power Mac G5 is a tad slower than the 3.2GHz Xeon.

CGTalk Moderation
01-19-2006, 12:00 AM
This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.