Cycles New Network Rendering on 640 Threads!

10 responses

1. asmithey

   9 September 2017 at 17:23

   Is network rendering functionality going to be part of Blender 2.8?

   Reply
2. Noc2

   9 September 2017 at 17:25

   Hi Odin, could you elaborate and share on how you went about accomplishing this? I’ll also second Asmithey’s question. Thanx

   Reply
   1. LordOdin

      9 September 2017 at 19:58

      Well I used our custom theory build which has the network render compiled in it since we paid for its development.. Though there are a few builds made by users since all the code is publicly available

      https://blenderartists.org/forum/showthread.php?437220-New-Cycles-Network-Rendering-Prototype

      Reply
      1. Juan Gea

         9 September 2017 at 23:56

         Hi LORDODIN

         Can you publish the economic part of it?

         I mean, what’s the cost of having those 10 machines working to render an animation?

         Let’s say for 36 hours to render an animation wich has a render time with all those machines of 30 secs/frame.

         That has been always my doubt about AWS.

         Also regarding this, does this network rendering system handle the distribution of the blend and dependencies like textures or alembic files?

         Cheers and thanks!

         Reply
         1. LordOdin

            10 September 2017 at 00:39

            To render for 36 hour it would cost roughly 576 dollars

            (The cost is 16 an hour for the 10 machines I selected 0.80 dollars an hour spot per machine)

            At 30 seconds a frame you can render 4320 frames in 36 hours… That is an entire short film

            36 hours of having the equivalent hardware at your house would cost roughly 21 dollars of power if your power costs 15 cents a kw hour.

            But you would also have to buy 10 20,000 dollar machines, install industrial grade ac (optional) and build an 8,000 dollar 20 gigabit lan.

            Sooo 576 dollars vs 208021 dollars lol

            Network rendering basically sends the ram of cycles in a separate blender instance after it has synchronized and built bvh of your scene.. And yes it has to send the ram to each individual machine so if your render takes 20 GB of ram you will have to send 200GB of files before you start rendering. Not at all optimal and I have stressed this to Lukas many many times

            Reply
            1. sambler

               10 September 2017 at 07:27

               While I agree uploading multiple copies of the data is bad, cycles isn’t the place to handle distributing the data between workers. And as you are rendering on your local machine, it should be the one responsible for distributing the data to each helper.

               Using AWS you should upload one copy and after starting multiple copies of blender, start rendering in one and it would then locally distribute the data to the others.

               While the netrender addon can do this, I am thinking a simpler cycles remote render addon could be made that will upload the data to one machine which would then distribute to it’s local machines.
            2. Juan Gea

               10 September 2017 at 13:20

               Hi LORDODIN

               Thanks for the answer, it’s oretty compeilling because it is a lot cheaper than renderfarms like Rebusfarm.

               Now how well does this scale if I want to reduce the time to a fourth of the time, let’s say I want to spend 2000€ in the render (wich I would never do being able to render 3 minutes of video in 36 hours! ) but is it just a matter of scaling computer power?

               I mean, if I hire 40 nodes instead of 10, do I get the time shortening? and is this time shortening in direct relation with the cost?

               Some kind of tutorial on how to setup all this is very welcome because I´m very interested on this, for some projects this could be awesome!.

               Thanks for all this and for your investment, I hope we can grow enough to make such an investment to and hire an inhouse developer to improve Blender :)

               Cheers!
3. Brian Lockett

   9 September 2017 at 19:05

   Oy, it’s so slooooow! Practically an eternity!

   ;)

   Reply
4. Reaction

   11 September 2017 at 09:08

   Watching those renders is indeed a thing of beauty! But I’ve never yet seen any network render solution that is beautiful, all of them getting bogged down in setting up clients/masters/slaves/schedules/IP addresses/folders, etc. I’m sure something many non-pro users want is just a simple (and I mean SIMPLE!) way to use all the computers in an office (i.e. on the same network) to overnight render the job running on one machine. In other words, on the Render tab, a single button that simply says “Slave mode”, which greys out all the other options and sets Blender listening for instructions. That’s it. On the manager PC, a little “Use network slaves” tick box, that’s it. Blender should sniff out all the slaves (even benchmark them first if necessary) but all I want to see is my animation frames appearing in a folder on my local PC, no setup required. I’m sure it’s possible; Blender’s existing network render was almost there already, but the laborious client/slave/master setup was never quite ‘cleaned up’ enough to make it easy to use (plus it usually returned black frames!).

   Reply
5. DaBorg

   4 October 2018 at 05:06

   I recently ran Blender’s classroom benchmark using a single ASUS B250 motherboard with a i7-7700K CPU and eight GTX1070 Ti GPU’s running blender v2.80 Alpha2. It’s nearly the same speed as this 640 core AWS instance.

   AWS instance with 640 cores: 40.93 seconds
   ASUS B250 motherboard with i7-7700K and eight Nvidia GTX1070Ti: 41.23 seconds

   Reply