Vray rendering on AMD vs Intel. Big problem

[ambarnair]

Hello,

Nice to meet all of you I have a strange problem that has been causing a fair amount of headache.

I recently purchased 5 HP Proliant DL165 1U rack mounted servers for the purpose of rendering. Each server has the following config:

 

2 x AMD Opteron 6274 2.4Ghz (24 cores total) with 64GB RAM (32 GB for each CPU, mounted in 2 pairs of 2 x 8GB each)

 

On my workstation where I fire off the renders, is an i7-960 with 3.3 Ghz and 24GB RAM.

 

A scene file using higher reflective materials with IRRmap (Medium-animation preset) and Lightcache (default settings) is taking 30 mins or so to render on the i7 at 1024x576. Sampling is DMC 2/8

 

My scenes are rendering 3-4 times SLOWER on the i7 than on the rackmounted servers! the same scene takes 1 hour 45 mins on the servers!

 

Just spent a decent amount of money on these servers, expecting better performance, but 3-4 times slower is just ridiculous. Granted the clock speed is more on the i7, but there are so many more cores on the servers.

 

Render on the i7 was a local render, and the server render is fired from backburner (both point to textures on the same storage server).

 

Is this a hardware problem? perhaps something in the BIOS?

 

OS is Win 7 pro on both machines. Thank you very much in advance

[ambarnair]

I also forgot to add, HP cannot help me with this problem. HP India has absolutely almost zero knowledge about any kind of 3d graphics, so they couldnt do anything about it.

[ambarnair]

I also just ran a cine bench test. For rendering the nodes score a 11.8 whereas the i7 that rendered the scene scored a 5.6. It's really strange

[Devin Johnston]

The AMD's are much faster than the i7, when the AMD's are rendering what's the CPU usage like?

[ambarnair]

Hi Devin,

Thanks for replying. Both machines show a 100% CPU utilization. I put a large evermotion set also up for render (a set with trees, so not much reflective materials) but last time I checked the i7 was still ahead while doing the irrmap calculation.

So I don't think it has anything to do with reflective materials specifically. I'm puzzled about the cinebench scores because the render tests on the amd system rendered pretty fast compared to the i7 but max 2011 renders super slow tried os reinstall but that didn't work either. Do you think it's some kind of bios setting issue? Thanks again

[ambarnair]

One more thing that may or may not help - the original scene I used as a test render is a scene with a car (reflections). With DMc sampler set to 1-16 it took an hour on the i7 and showed 3.5-4 hours on the amd. When the sampling was set to 2-8 I got 30 mins vs 90-120 mins respectively so it seems that the renders are slower on the amd by a scale of 3 or 4.

[Devin Johnston]

How are you rendering to these slaves, I mean are you doing these texts through DR or just sending them individual scenes? I assume all the GI settings are the same right?

[Francisco Penaloza]

How is you Hdd configuration, you are reading the textures from a separate computer?, The optimum set up should be, a server for textures, proxies etc, a separate server or Hdd for Final images, a work station, render nodes.

If you have your render nodes reading and writing textures at in the same hard drive, they will be s little slower but not that much slower.

Everything is in a wired networking right? they are all connected to the same hub?

I am assuming all network cards a gigabits.

 

If they are rendering, reflection or flat proxies it should be the same, if it is slow is something else not because the reflection of your materials.

 

if you run cinebench that's is a local rendering, so that's why I guessing that your problem could be on the network.

[ambarnair]

Hi devin, Francisco,

The render on the i7 as well as the rendernode was through back burner. Both nodes are on the same network (connected via hp gigabit switches), the scenes assigned were exactly the same scenes. All textures, vray proxies etc are being pulled from the same file server. Ie the same max file was sent to both machines. The output file is an openexr with all render elements that is being written to the same server but both nodes are doing this and rendering the exact max file (we tested at different times as well but the problem persists ). Did I make an (expensive) mistake going with amd instead of intel?

[ambarnair]

Also ran a test with an evermotion tree with no leaves multiplied 100 times and ran a max 2011 mental ray occlusion render. I7 shows 7:35 and the amd shows 8:35. No textures used

Do you think it is a CPU core to ram ratio?

[ambarnair]

Doh! I'm so sorry I made a typo in the processor - it's a dual amd opteron 6234 not 6274 but still with 12 cores/CPU and 2.4ghz. Latest update - I installed a graphics card of 1gb in the amd node and the rendering has improved - now shows 1hr render time for something that renders in 30 mins on the i7. So it's improved but getting a gpu for each rendernode is a bit of a pain but still it's slow..

[matthewgriswold]

Adding GPUs with lots of RAM will do nothing for your rendering speed unless you are rendering with GPUs.

GPU RAM is not equal to DDR3 RAM. CPUs utilize the DDR3 from the server board.

 

I think we need to see some rendering settings.

[ambarnair]

Hi Matt,

I'm out of the studio so can't post render setting screenshots but I can say everything is set to default except adaptive DMc sampler with 2/8 and gi set to irradiance map (medium preset) and light cache (default). Irrmap is set to single frame and default for single frame render for light cache.

Have tried brute force with 10/6 but no difference. Hope this helps. Environment and reflective override are unchecked.