Forum Groups
  All forums
    Help & Feedback
      Work in progress
      Finished Art
      Non-Max related

Maxunderground news unavailable

Intel Xeon Phi Experience
show user profile  Krogoth
I was asked to look into an in-house render farm soloution, and I came across the Xeon Phi Coprocessor cards. Has anyone had any experience with these, and do you know if they work with 3ds Max and Mental Ray?

In case you don't know, the coprocessor card has 57 to 60 processors on them, depending on the model:

I have also heard that Max can only access up to 36 cores.

read 1177 times
8/22/2015 7:50:24 PM (last edit: 8/22/2015 7:50:24 PM)
show user profile  Mr_Stabby
Despite being based on x86, to the card is more similar to a gpu rather than a cpu meaning that they won't be extra cores you see in task manager but have to be specifically invoked through a specialized api.

They are OpenCL capable so I guess you could run vrayRT on them all though, all the gpgpu renderers available are optimized around the SIMT model rather than MIMD.

SIMT - single instruction multiple threads - is what gpus use across thread blocks. 100% efficient when there is no branching but that drops significantly if there is branching. Because all SIMT threads run in a sort of lockstep, if even one thread branches somewhere, all the others must wait for it to finish before continuing.

MIMD - multiple instruction multiple threads - is what the Phi is using and is basically like the different cpu cores, unimpeded by what others are doing.

among other differences not attributed to the thread model but rather to gpgpu vs phi, the phi has lower total bandwidth, lower latency random access, faster cross thread operation and lower peak performance per transistor(or dollar).

So while code optimized for MIMD being forced to run in a SIMT environment would be much worse than vice versa, the commercially available renderers arent able to use the strong points of the Phi because its kind of an unique thing that they aren't optimized for and because of that, get run over by the raw throughput of gpu's.

Now on a personal note, I'm somewhat disappointed in the phi. It came too little, too late and the fully implied amp support that they were touting before release never materialized. It does have its niche uses but keeping its current ratio of flops per dollar, it still cant compete in rendering or realtime physics or most other simple tasks(yes those 2 are computationally relatively simple) with the gpus even if those algorithms were optimized for them. Considering that and the fact that the card is kind of an oddity with future development depending on intel's whims, most companies would rather not invest their coding time into creating software for it for other than special cases.

That said, I still hope it takes off since coding for MIMD is way more comfortable than coding for SIMT.

read 1159 times
8/22/2015 9:05:52 PM (last edit: 8/22/2015 9:09:18 PM)
show user profile  Krogoth
Thanks for the trove of information, it was very helpful.
read 1134 times
8/23/2015 1:59:40 AM (last edit: 8/23/2015 1:59:40 AM)
show user profile  ScotlandDave
Stabby i love the fact that you haven't just copy/pasted that from Wikipedia, that you actually KNOW that stuff..

In saying that - do you keep up with computing tech at that level out of personal interest or is it related to work you do?

Website | Blog | Contact | Vimeo

read 1132 times
8/23/2015 2:25:47 AM (last edit: 8/23/2015 2:25:47 AM)
show user profile  msimecek
i'm using an e5-2699 2.3 ghz. 18c/36t

it's quite fast especially when rendering in max but it doesn't replace an actual render farm.

try a money vs. render power approach... setup up your host as this high end chip, then get a few machines , still powerful but yesterdays news, and hook them up with back burner.

but... max 2016 using backburner has a bug that renders real slow using it. 6x to 10x slower. so if you go this route use 2015 or earlier.
read 1131 times
8/23/2015 2:53:17 AM (last edit: 8/23/2015 2:53:17 AM)
show user profile  Bolteon
Just use the cloud.

-Marko Mandaric

read 1126 times
8/23/2015 3:44:13 AM (last edit: 8/23/2015 3:44:13 AM)
show user profile  Krogoth
Yeah, I was trying to find where the line is for more cores at lower clock speeds vs less cores at a higher clock speed, which is why I was looking at the Phis. I was hoping the Phis would be an option and then we could just put vSphere on and run some VMs.
I have not noticed the issue with 2016 BackBurner, I'll keep an eye for that, as we do use 2016.

We use RealFlow and we calculate out FG maps, and we don't want to have to deal with transfering all of that to a cloud service. Though I did throw out RebusFarms as an option to the person who asked me to look into render farms.
read 1117 times
8/23/2015 5:49:45 AM (last edit: 8/23/2015 5:49:45 AM)
show user profile  Mr_Stabby
>do you keep up with computing tech at that level out of personal interest or is it related to work you do?


read 1105 times
8/23/2015 7:04:59 AM (last edit: 8/23/2015 7:04:59 AM)
show user profile  Bolteon
"We use RealFlow and we calculate out FG maps, and we don't want to have to deal with transfering all of that to a cloud service."

So calculate those in the Cloud, and then render with them up there.



One way or another, having faster internet is a lot less work to make happen than dealing with computer hardware and software.

-Marko Mandaric

read 1097 times
8/23/2015 10:09:11 AM (last edit: 8/23/2015 10:09:53 AM)
show user profile  digs
we have some xeon blade servers at work - each one is 32 core, we use max, c4d, vray - fluid caching, rendering, etc.

they seem pretty quick
read 1047 times
8/24/2015 2:12:01 PM (last edit: 8/24/2015 2:12:01 PM)
#Maxforums IRC
Open chat window