iray benchmark

Become a member of the CGSociety

Connect, Share, and Learn with our Large Growing CG Art Community. It's Free!

THREAD CLOSED
 
Thread Tools Search this Thread Display Modes
Old 12 December 2010   #46
Originally Posted by ClausD: My second GTX 580 arrived this morning so I switched the last one of the GTX 285 with it.

100 Interations in 16 seconds

10 minutes result on my Quadcore 6850 Extrem @ 3.0 GHz with 2 x GTX580




don't get this wrong but that is a bit disappointing, I was expecting less noise in 10 minutes in such a simple scenario. I mean, it's 2 GTX 580...
 
Old 12 December 2010   #47
Lightbulb Xeon 4.2 GHZ + GTX 460

Hi!

Here are my times.

And I've got idea: Maybe we should also post GPU times only, because CPU's may "distort" the times of GPU - and this thread may help with choosing GPU best for iRay?

So, here are my times, three options (i've turned on/off GPU/CPU use with this script):

1. Intel Xeon @ 4.2 GHz + GAINWARD GeForce GTX460 2 GB RAM:

100 passes: 0:35 (35 seconds) - translation 0:03

10 min image:




2. Only GAINWARD GeForce GTX460 2 GB RAM:

100 passes: 0:46 - translation: 0:02

10 min image:




3. Only Intel Xeon @ 4.2 GHz:

100 passes: 1:33 - translation 0:06

10 min image:




So, definitely GPU does the job ;]


Does anybody has TESLA C1060 or Quadro FX 5800?

I'm curious about performance of these compared to GTX460....
__________________
Visit my iRay 3D Portfolio:
www.rendergarden.org

Last edited by kcpr : 12 December 2010 at 01:22 AM.
 
Old 12 December 2010   #48
Originally Posted by CHRiTTeR: How does vrayrt perform with both gpu and cpu on? One would suspect it would be faster? or is there some bottlenecking going on?
You can run it this way, but there is little point in it; with the GPU being 20+ times faster, the CPU plays an insignificant part in the result.

On the other hand, V-Ray RT on CPU has way more capabilities that the GPU version (and iray too), including dynamic geometry like proxies with true instancing, layered materials, textured lights (other than a dome light, that is) and nested procedural shaders.

Best regards,
Vlado
 
Old 01 January 2011   #49
Just finished building my new workstation, and so glad to finally be able to post this here.

CPU - i7950
GPU - 580GTX

100 Iterations - 23 Seconds

10 Minute Render

__________________
www.psvisuals.com - 3D Visualization and Content Creation
 
Old 01 January 2011   #50
Originally Posted by metamesh: don't get this wrong but that is a bit disappointing, I was expecting less noise in 10 minutes in such a simple scenario. I mean, it's 2 GTX 580...

Not as disappointing as 4x480 performing only 2x compared to a single 480. Supposed all this gpu thingy was way more scalable, e.g. 4x 480 doing 3.5x a single one at least.

EDIT I could've swore I read 22s on a single 480. Instead, it's 28s on a 470, vs 11 on 4x480.
a bit of (probably wrong) math: based on the cuda cores, the theoretical time for a single 480 should be around 26s. so 11s for 4x480=2.36 faster than a single one. Still not that good imho...
Of course, since iray uses both cpus and gpus, every calc is bias
__________________
My vimeo stuff

Last edited by simonenastasi : 01 January 2011 at 01:31 PM.
 
Old 01 January 2011   #51
Originally Posted by simonenastasi: Of course, since iray uses both cpus and gpus, every calc is bias


That's why I suggested to use only GPU Benchmark, using iRay Manager...
__________________
Visit my iRay 3D Portfolio:
www.rendergarden.org
 
Old 01 January 2011   #52
Originally Posted by metamesh: don't get this wrong but that is a bit disappointing, I was expecting less noise in 10 minutes in such a simple scenario. I mean, it's 2 GTX 580...
Originally Posted by simonenastasi: Not as disappointing as 4x480 performing only 2x compared to a single 480. Supposed all this gpu thingy was way more scalable, e.g. 4x 480 doing 3.5x a single one at least.
I know it's frustrating. This Brute Force via GPU fashion is something that is good for users, but really really great for hardware vendors.

The reason for this to happen is that the quality vs GPU power in brute force rendering is logarithmic. It's similar to how our eyes perceive light. Where doubling the light intensity doesn't mean double perceived intesity. An easier way to think about this is: Imagine a rolling snow ball growing as it rolls. The more it grows, more and more snow is needed to make it bigger, up to a point that for doubling it's size might need huge amounts of snow.

With IRay and any other progressive renderer it's just the same thing. The more time passes, the amount of samples per second the hardware is able to pull off (that is constant) will be less and less effective. So you might need 4x more cards to say double the quality. But if that quality is still not enough, you might need 20x your previous count to double quality again.

You can test this rule with rendering time. Let a render cook for 10min, then try to wait for a render until it reaches around twices less noise as the 10min one. It probably will take much longer than 20 min.

Hope this makes sense...

Regards,
Jr.
 
Old 01 January 2011   #53
Originally Posted by Eugenio: I know it's frustrating. This Brute Force via GPU fashion is something that is good for users, but really really great for hardware vendors.

The reason for this to happen is that the quality vs GPU power in brute force rendering is logarithmic. It's similar to how our eyes perceive light. Where doubling the light intensity doesn't mean double perceived intesity. An easier way to think about this is: Imagine a rolling snow ball growing as it rolls. The more it grows, more and more snow is needed to make it bigger, up to a point that for doubling it's size might need huge amounts of snow.

With IRay and any other progressive renderer it's just the same thing. The more time passes, the amount of samples per second the hardware is able to pull off (that is constant) will be less and less effective. So you might need 4x more cards to say double the quality. But if that quality is still not enough, you might need 20x your previous count to double quality again.

You can test this rule with rendering time. Let a render cook for 10min, then try to wait for a render until it reaches around twices less noise as the 10min one. It probably will take much longer than 20 min.

Hope this makes sense...

Regards,
Jr.

Eugenio, is a bit puzzling to me, 'cause the comparison was made on a fixed value (100 iterations), not based on quality.
__________________
My vimeo stuff
 
Old 01 January 2011   #54
Could someone please provide a *.obj or *.fbx file instead of *.max?
I would love to run this scene in octane render, also. And Im no max user.

Many thanks in advance!
-c-
 
Old 01 January 2011   #55
Originally Posted by cornel: Could someone please provide a *.obj or *.fbx file instead of *.max?
I would love to run this scene in octane render, also. And Im no max user.

Many thanks in advance!
-c-


http://forums.cgsociety.org/showpos...51&postcount=19




and here is another test result with my new CPU (Intel i7 2600 3.4Ghz)

1. 10min render

2. 100passes rendertime : 1m40sec
3. spec - CPU : i7 2600 (3.4Ghz no over), GPU : no cuda
 
Old 01 January 2011   #56
Some interesting stuff I found running some test, like it is faster to run GPU only (mental_ray_string_options.addoption "iray threads" 0) than GPU+CPU. Here are the stats from the 10 min benchmark:
Originally Posted by CPU Only: RC 0.3 progr: progressive: finished (sub)frame 594 in 1.10562 seconds
RC 0.3 info : progressive: stopping rendering due to time constraints
RC 0.3 info : progressive: total rendering time was 599.72 seconds
RC 0.3 info : progressive: average frame time was 1.01 seconds
Originally Posted by GPU and CPU: RC 0.3 progr: progressive: finished (sub)frame 3221 in 0.75172 seconds
RC 0.3 info : progressive: stopping rendering due to time constraints
RC 0.3 info : progressive: total rendering time was 599.72 seconds
RC 0.3 info : progressive: average frame time was 0.74 seconds
Originally Posted by GPU Only: RC 0.3 progr: progressive: finished (sub)frame 3525 in 0.68200 seconds
RC 0.3 info : progressive: stopping rendering due to time constraints
RC 0.3 info : progressive: total rendering time was 599.33 seconds
RC 0.3 info : progressive: average frame time was 0.67 seconds
Which shows 1.09438 speed increase just using GPU over GPU+CPU, and 5.93434 speed increase of GPU over CPU. 100 Iterations Benchmark:
Originally Posted by CPU Only: RC 0.3 progr: progressive: finished (sub)frame 100 in 0.97438 seconds
RC 0.3 info : progressive: stopping rendering due to max frame constraint
RC 0.3 info : progressive: total rendering time was 78.05 seconds
RC 0.3 info : progressive: average frame time was 0.78 seconds
Originally Posted by GPU and CPU: RC 0.3 progr: progressive: finished (sub)frame 100 in 0.69917 seconds
RC 0.3 info : progressive: stopping rendering due to max frame constraint
RC 0.3 info : progressive: total rendering time was 18.00 seconds
RC 0.3 info : progressive: average frame time was 0.55 seconds
Originally Posted by GPU Only: RC 0.3 progr: progressive: finished (sub)frame 100 in 0.64087 seconds
RC 0.3 info : progressive: stopping rendering due to max frame constraint
RC 0.3 info : progressive: total rendering time was 15.57 seconds
RC 0.3 info : progressive: average frame time was 0.47 seconds
System:
Dual Xeon 5620
SuperMicro X8DTH-6 Motherboard
24GB Ram
Tesla C2050 (x8)
Quadro 5000 (x8)
Quadro 5000 (x8)

I am sure there would be some more speed differences if the motherboard supported x16 PCI-e lanes, but unfortunately I had issues with Dual Xeon motherboards with multiple x16 slots. No picture since it is too large to attach here.

-Eric
__________________
"The Evil Monkey hiding in your closet."
 
Old 01 January 2011   #57
This is interesting to point out, and I think it varies from scene to scene. In some of my scenes, for example, enabling my GPU doubles my speed. In this specific benchmark scene, it is over 10 times faster. So while I think the GPU helps a lot, it isn't a guaranteed even boost in all scenarios.
__________________
www.psvisuals.com - 3D Visualization and Content Creation
 
Old 03 March 2011   #58
10 minute render image included (hopefully)

100 iterations in 16 seconds with just the GPUs

100 iterations 18 seconds with both CPUs and GPUs

Corei7 the entry level one
24 gigs of ram 1600mhz
quadro fx 5600
gtx580
gtx280

Its really tough to get rid of all the noise huh?
Attached Images
File Type: jpg RenderTest2.jpg (73.9 KB, 58 views)
 
Old 03 March 2011   #59
Dual XEON E5520 @ 2.27GHz (core i7s, 16 virtual cores)
Quadro FX 5000
24GB RAM

When I rendered GPU only, the render time for 100 iterations went up significantly. In my case, the CPUs play a significant role. PiXeL_MoNKeY, in your case, I can see that with 24 badass GPUs, the CPU is probably just getting in the way.

We also see the logarithmic scale of cost vs speed. I have one Quadro 5000 and my render time is only twice that of yours... or did you render at higher resolution?

Regardless, I'm envious of your GPU farm. Is that a shared resource in your studio? Tell us more about your setup.

100 iterations: 31s
10 min render:


EDIT:
The quality and render time of my rather expensive rig is almost identical to the OPs results at spec - CPU : i7 860 (3.4oc), GPU : GTX460. His was 33s. I copied his 10min image and compared it to mine in PS and they are nearly identical.
__________________


Last edited by jonahhawk : 03 March 2011 at 06:21 PM.
 
Old 03 March 2011   #60
core i7 950 (8 virtual cores) 3GHz 6GB tri channel DDR3
GeForce 8800 GT 1GB

100 iterations: 1:22

__________________

 
Thread Closed share thread



Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

vB code is On
Smilies are On
[IMG] code is On
HTML code is Off
CGSociety
Society of Digital Artists
www.cgsociety.org

Powered by vBulletin
Copyright 2000 - 2006,
Jelsoft Enterprises Ltd.
Minimize Ads
Forum Jump
Miscellaneous

All times are GMT. The time now is 10:57 PM.


Powered by vBulletin
Copyright ©2000 - 2017, Jelsoft Enterprises Ltd.