Raspberry Pi Cluster Blender Rendering Farm

Thought I would put together a cluster build that might actually be useful, we are building a raspberry pi cluster blender rendering farm!!
○○○ LINKS ○○○
Raspberry Pi 3 ► amzn.to/2rrixtm
tutor4u ► / tutor4u
○○○ SHOP ○○○
Novaspirit Shop ► goo.gl/NraX9c
Amazon Store ► amzn.to/2AYs3dI
○○○ SUPPORT ○○○
patreon ► goo.gl/xpgbzB
○○○ SOCIAL ○○○
novaspirit tv ► goo.gl/uokXYr
twitter ► / novaspirittech
discord chat ► / discord
FB Group Novaspirit ► / novas. .
○○○ Send Me Stuff ○○○
Don Hui
PO BOX 765
Farmingville, NY 11738
○○○ Music ○○○
From Epidemic Sounds
patreon @ / novaspirittech
Tweet me: @ / novaspirittech
facebook: @ / novaspirittech
Instagram @ / novaspirittech
DISCLAIMER: This video and description contains affiliate links, which means that if you click on one of the product links, I’ll receive a small commission.

[@NovaspiritTech]

come join me on twitter! twitter.com/NovaspiritTech

[@pdjames1729]

gonna take #risc5 #sifive Broken-source a Decade to get anywhere near this. Then #ARM will be (even more) Open and #BlueGroup will have failed. Like Me :D :nP

[@pdjames1729]

Nova, is expansiveness that's the big deal here. Once you're setup - you can add 6ghz for $50 ish... indefinitely... why don't you? ;o) #FastestPiRenderfarmOnEarth ;) or, period :D

[@srfrg9707]

Novaspirit Tech Hi. I used to work as an animator for several studios out there. The concept of a render farm is not only about parallel processing, it's mainly about time sharing and job automation managment with priority of production. In other words a propre render farm known which job it has to render first based on production priorities and stuff. A typical animation movie hires sometimes hundreds of animatirs and rendermen. The render queue is changing all the time as all the scenes are delivered for rendering and it's not efficient to render them in the order they are submitted. The priority is then defined by algorithms.

[@bogas78]

Good info I never made cluster from rpi's as for me there wasn't real life use but this is really good news as I got around 15 rpi3 laying around so now they can be used to speed up some things :)

[@rolandbjorn7982]

dont know if anyone gives a shit but if you're stoned like me during the covid times you can stream all of the new movies on Instaflixxer. Have been binge watching with my gf lately =)

[@TerabitTech]

It would be nice to compare a 30RPi cluster to a 900$ rendering PC to see which one performs the best

[@NovaspiritTech]

i would love to see that too

[@st3ddyman]

I’d be very surprised if the PC wasn’t a lot quicker

[@crazed357]

Knowing I build my pc for around that price, it’s hard to imagine even 30 rpis would render 3D faster or even half the speed.

[@alexandracrawford800]

On the other hand it depends on the size of the chunk each core is supposed render, with optimised chunk size, you may be very, very surprised especially with GPU concurrency, there's a lot that has not been dealt with by the corporate thinkers. Why work at a problem when you can throw a ton of money at it! This is the time when a new industry could be born just as gamers started back in the 80's with Render Plus Interest auto render farms, auto billing? Why not? Media is BIG business!!!!

[@crazed357]

Alexandra Crawford Well the industry is moving towards the ARM architecture. Windows now can run on arm, Apple spent billions into having the fast biggest cached Arm CPUs, Amazon is phasing out their intel servers with their own custom ARM servers. I do believe there is a place for raspberry pi clusters in the future. I would gladly pay $50 extra for the raspberry pi foundation to add A72 cores for their PIs to pave way for the future of ARM based desktops/Soc builds

[@hugoortega195]

After watching multiple videos about Raspberry Pi cluster finally someone actually showed a use case for this. Thank you.

[@tibfulv]

You could presumably also use it for a web browser, or a database.

[@slave288]

I absolutely agree, clusters have always looked very cool to me but had very little real world use in my opinion. For a while now I've ben wondering about making a cluster to do Video Rendering using Kadenlive but didn't think it was possible. I'm glad to see that this is on the verge of becoming possible!

[@alexandracrawford800]

..AND NOW THE AVAILABILITY OF CM4 MODULES (2-8GB, 32GB EEPROM) + CLUSTER BACKPLANES WITH PCIE SUPPORT FOR HIGH SPEED NETWORK SHOULD PROVIDE AN AFFORDABLE, LOW POWER SUPER COMPUTING CLUSTER FOR ANYBODY WHO WANTS TO TRY DEVELOPING SOME INTELLIGENT MULTI-CORE TASK ALLOCATION... who says it's all been done before? Sounds like Au (Artificial unintelligence)

[@dgreene2701]

It is probably useless for what he is trying to do. But a pi cluster is a pretty advanced network tool. It mostly useful for network security, server ops, or database management. If you don't deal with those then it's probably useless to you.

[@algotrash]

The main problem is ram quantity on the rpi. It works here because there is no textures to load in memory. Still a cool proof of concept.

[@NovaspiritTech]

i did the BMW stress test on the cluster and was only using 180~mb ram but yes once you get into animation with more textures, ram might not be enought.

[@techwizpc4484]

@@NovaspiritTech You could try SDCs equipped with higher RAM like Odroids.

[@tibfulv]

Or rendering pictures in segments.

[@seditt5146]

@@techwizpc4484 The Rockpro64 I just got has 6 cores and 4gb of ram, a little more expensive then smaller boards but surely would run something like a render farm smooth as butter if a few were setup together.

[@techwizpc4484]

@@seditt5146 Or perhaps for those stuck with pi clusters, send the images in tiles rather than whole frames so that it's easier for the pi's to render. I wonder if that can be done.

[@sebastianriveros3046]

I had a little discussion a month ago about how well a rasperry cluster could work. This is a very nice proof of concept, and I would really like that someone made like 20 or 30 RPi running against the best CPU you can buy for the same money. As you said, there has been this transition in the last years in favour of ARM servers. I'm super glad with this and I really hope that in a few years, we can stack like 10 RPi running blender to speed up rendering times. Thank you!

[@zachburke8906]

Sebastian Riveros well here’s a 40 node build. kzbin.info/www/bejne/n6GxdYV8ls16lZI I also know a few 100-500 node raspberry pi builds exist. The problem is for running large blender farms will be the ram, for simple things like shown in the video it’s not a really a problem. As soon as it hits anything complex it will be maxing out the ram and slowing down a lot.

[@sebastianriveros3046]

@@zachburke8906 Super awesome! thanks! I'll check it out!

[@spydergs07]

With the Pi's, it took 7 mins. Without the Pi's it took 19 mins. That's actually pretty impressive.

[@juweinert]

Proof of concept? Actually you presented real world value. I mean after all your "quick and dirty" farm did better than your machine... So if you'd use your laptop as node too you could use the little farm as render-acceleration. Pretty cool stuff, I was really surprised these 4 SBCs did that well!

[@alexradu1921]

yes i thought that too.. why wouldn't use the laptop too?

[@mark8664]

It is soooo easy to follow your video because you talk like you are having a conversation with us, really great!

[@DrNickDoom]

Can you make the Custer run the blender benchmark

[@blendpinexus1416]

i think the blend file would take a lot of modification to be able to split across different devices, or a special software will be needed to do the split for you.

[@yetidynamics]

cool. I toyed with this Idea awhile ago, but the lack of ram made it non practical... however!, your video encouraged me to look around, and it looks like there are some alternatives to the Pi now that might make this worthwhile for me

[@shoaib2307]

Well you can try it now

[@yetidynamics]

@@shoaib2307 nope. AMD came out with insane multicore cpus for cheap prices, so i went with that, more cost effective, and faster

[@yuan.pingchen3056]

@@yetidynamics which amd processor can beats nvidia gpu on rendering?

[@yetidynamics]

@@yuan.pingchen3056 All of them, Because my Nvidia GPU fails to render anything when it runs out of RAM.. My scenes very typically go into the 40GB range of memory usage.

[@autonomous2010]

@@yetidynamics Large scale Fur, Water, Fire, etc simulations pretty much make GPU Rendering a No-Go currently.

[@teitoklien]

I don’t get why this guy has so few subscribers he’s fucking awesome and actually shows real stuff

[@williamhart4896]

Interesting Don have at one point thought of building a arm cluster computer nice content

[@alexanderwingeskog758]

Did sort of something similar but about 26 years ago... With Lightwave and Screamernet... It was super easy to setup, it probably had half of a half of one RPI2 as master (my computer ofc) and then about 8 computers (from 133 Mhz AMD, to 200 Mhz Intel processors) all running Win95 (as I remember, maybe it was Win98). Anyway... it was super simple, and it was fast... It did send the every resource needed (textures and what not) to every node, then the master just send "render this frame" when told every other node to skip that frame and when the that computer was done it got sent another one and so on... You did not need any setup on the clients... it was just "install this" and put you computer on... Was awesome... (about 26 years ago)... I made lots of trial and error animations... I had almost 10 times the power of my "master" computer.... I think with a blender plugin that could use the GPU on the RPI and the CPU's this could be awesome... and if you could install everything as just a "node" in pretty much everything you own (maybe with a "power slider", like a phone maybe you don't want it do drain the batteries too much) you could just add your old tech to make the "local" renderfarm bigger/better/faster... And if you are not happy buy resources online.... But hey... I got probably 8 phones that are almost up to scratch and with proper usage of the power could easily contribute to a frame of two :-) and about 8 computers... With multiple CPU's and GPU's... But it has to be easy setup... install a node, install a "master" and let master control nodes easily... make use of old shit :-)

[@kungfujesus06]

Select GPU for the openCL renderer in Cycles and I'll bet it smokes the Pis. The memory bandwidth is actually a pretty big issue for the Pis. Here is a use case where the 100mbit networking of the Pis doesn't kill you (mostly because the work to render is finishing far slower than 100mbps, even if the Pis were rendering the frames at a less than crappy rate).

[@BanditLeader]

Just download more bandwidth

[@sanantoplic8715]

It is cost effective for small projects i.e. Furniture Renderings thanks for the inspiring video

[@dgreene2701]

It's a fun experiment in order to learn about cluster computing, if that's the things that gets you interested in the concept. But I would be fairly confident that even basic GPUs could out perform that cluster farm in rendering Blender scenes. CPUs just aren't designed to do multi-threading. However, the same clustering concept can be applied to GPUs. They are very common for this application as well as for any kind of AI.

[@TheStevoth]

That's very useful. I didn't know that blender could be used like that. I'll put all my old laptops to use (and my raspberry pi(e)s),

[@dainipeagram4837]

This is exactly what I was looking for. For setting up blender on several servers, all running VM's

[@SlowPCGaming1]

You may not see this but I think it is time to tackle sci-fi voice commands like on Star Trek so that we tell the computer cluster what we want it to render and it does the rest. Then we don't have to bother with fiddling with the commands of each program, or how to configure the hardware, stuff that makes the average person roll their eyes and lose interest. Is that something you can achieve using these types of computers or more powerful Ryzen based SBCs?

[@DevinFlake]

Very cool project! I really like projects with real world application :)

[@TerabitTech]

I have many PCs around the house as me and my parents work with PCs. Could I do a rendering farm with Live Linux distros, although the PCs are quite different from each other? I could use at least 3 laptops, two desktops, an Odroid XU4 and a Raspberry Pi (1st gen)

[@NovaspiritTech]

yes, as long as you can install blender, you should be able to network render

[@lilprincessuwu9292]

It will work with almost anything. I would recommend throttling it if you use laptops since there not really designed for running max processor all the time.

[@tibfulv]

Indeed, this sort of configuration predates affordable SBCs. In the early noughties it was used on standard PCs, and was known as a Beowulf cluster. And people really hooked up almost anything to it. DEC Alphas, 386s, Pentiums, you name it.

[@debconfsetup8673]

simple answer: yes, you can! more complex answer - you should not. laptop and android devices usually have less compute power, less memory. you should keep updated blender version because newer shaders don't compute on older versions. and provocative: 32bit or 64bit? sheepit-renderfarm.com refused 32bit blender render because result images have slightly different colors.

[@kungfujesus06]

While the problem's distribution makes inter node communication less of a bottleneck, you'll still be limited by the slowest node in the render farm. E.g. you'll have a node chewing on one frame for large durations of time where a different node could have done 5 frames in the same amount of time. This is a problem of course that is pretty generic to distributed computing.

[@murphyking79]

This is the best cluster video I have seen!

[@lime-407]

Can you make an Video how to use more than 2gb of ram ? Pleeeeeeeeeese 🙏

[@TheBinklemNetwork]

awesome video! you give me tons of topics to talk to my dad about, something that I've very much needed!

[@frroossst4267]

I have a Centrino laptop and I'm gonna try this setup cause I too made the planets collision animation took 1.5 days and I'm really hoping this would speed it up. Thanks! Cheers, Adhyan

[@blendpinexus1416]

ooo, i think having a mixed cluster will be more informative, it sure was for me.

[@Valnurat]

Do you think you can update this video with the latest version of blender and the option to do network blender on RPi's?

[@Ckhpo]

Cool proof on concept. I'm researching render farms for a forthcoming animation project. This was great insight for how we might use the Raspberry Pi.

[@delvorin1841]

Hey Don, I have been a subscriber for a long time and this I believe this is my first post. I would recommend redoing the render on your pc with but this time enables GPU rendering (If that was not the case.) Of course, this is dependent on how many Cuda cores your GPU has.

[@NovaspiritTech]

to be honest, my cpu ryzen 1700 compiles faster then my nvidia 1070 gpu by 20%.. but yes this was compairing cpu to cpu not cpu to gpu, i will do a bigger render compairison and upload to my twitter account

[@delvorin1841]

@@NovaspiritTech You are correct about being cup vs CPU, not GPU. btw I only ever made one youtube video and it was a blender animation based on a tutorial from Tuts4u. Here is my grand animation....lol kzbin.info/www/bejne/f6qmnYd-fZehpZo

[@NovaspiritTech]

@@delvorin1841 that's awesome!!n

[@keltonjohnson6197]

@@NovaspiritTech That sounds a little off to me. I checked in Blenchmark and the GTX 1070 should be about 5% faster at rendering than a Ryzen 1700 (maybe part of the 25% difference is an awesome overclock?). By chance, did you leave your tile size at the default setting when using the GPU? I say this because if you left your Tile Size at the default, it will be set at the default size for a CPU thread which is 64x64, and that is crippling for a GPU. GPU optimized tile size is more like 256x256. And CPU-optimized size can even be smaller like 16x16. I like the tile size table in this article. www.blenderguru.com/articles/4-easy-ways-to-speed-up-cycles.

[@NovaspiritTech]

@@keltonjohnson6197 lol like I said I'm noob to blender. I didnt change the tile size for gpu

[@TriVoxel]

That flickering from your screen recorder is likely due to a faulty Intel graphics driver called xf86-video-intel. I'd look around for how do disable that driver if I were you as Intel drivers are now in the kernel anyways!

[@kebman]

This was awesome! Thank you! Not to compute the answer to the universe and everything...

[@BlenderBeanie]

Ayyy this planet tutorial was pretty much the first blender tutorial I watched! Good nostalgic feel, seeing it after many years! The only reason I clicked on the video tbh :D

[@benricok]

Waited so long for a video about this subject!!!! Awesome💥💥

[@thomasstorbugt6829]

What Operating system are you running on your laptop in this video? Really cool video btw, makes clusters even more interesting to build

[@davidmurphy563]

I wouldn't use that method personally. I would run via the gui in each unit, save as a placeholder png to a network location. That way each unit goes straight to a new file the moment it's ready to. I suppose headerless is an advantage with your approach but with mine you can add and remove devices as you go and I'd presume it's less likely to drop frames / fall over. The real problem is running a ram heavy textured, high poli png on a rpi. It might so slow there's little point. Could you maybe test a 1 min 30 sec realistic frame?

[@gormanspacemarine]

God Skill. That was frigging awesome. MORE!!!

[@keithmiller9665]

I do quite a bit of ffmpeg transcoding of movies and I was wondering if this could be done on a raspberry pi cluster?

[@chetana9802]

2 years seems right time for a 2nd run

[@coolduder1001]

Would an old server be cheaper at this point? I mean it's basically silent and doesn't use as much power but how well would it work before the price gets to high. I'd probably use some orange pi or pi zeros

[@GameWorkflow]

Awesome video! I need to try this! (Also look into blender's denoising feature for that planet collision of yours, its really useful for getting rid of that pesky noise and fireflies)

[@RinksRides]

so $150 of RPi 3's are 3X more powerful than i7 ~$150 laptop? I have a bout 5 RPi 3's and an old E6410 with a quad core HT i7-740Q 4C8T i believe, also have a VM Win 10 Workstation with 16 threads assigned and 24GB CPU passthrough via QEMU under unRAID. Would be interesting to do a head to head to head with a 16 node R Pi 4 cluster. I also have 2x Atomic Pi's I'll try this on.

[@bacchiachar754]

Can we use a mining rig to GPU render in blender.(it make it more compact ).i never tested this mining motherboards can actually run blender . It would be very helpful.🙏

[@anandmoon5701]

Super cool stuff, this bring new ideas to work on SBC.

[@TwinTailTerror]

can you update this? id like to see if the pi zero / zero w can be added i have some "broken" pi and i want to put them to use rather than throw them away. most have a few broken pegs or one has missing disk that the "power" would plug in. so cant use hats i have no use for them and such. anywho im hoping you can help or maybe update this i really wanna use these pi

[@sennabullet]

awesome concept, awesone vid...thank you so much!! Request...pls zoom in on your screens when you type commands and maybe a bigger more visible mouse pointer. cheers!!

[@operator8014]

Damn... Not bad for a dual core CPU. That's almost as good as my old Q6600...

[@ryanwakebradtelle8682]

Any idea how big the biggest pie cluster is, I mean if there's 7 pie servers at different universities each outclassing the last how am I supposed to know if another newer server has 50,000 more pies?

[@lawrencedoliveiro9104]

1:52 2.78 -- that’s ancient in Blender terms. ;) Blender is such a fast-moving target -- the current master branch version is called 2.79.6, with 2.8 already in beta. I was going to say download a build, but blender.org does not provide ARM builds. I build my own from source, using a wrapper script to set all the CMake options I normally use.

[@lawrencedoliveiro9104]

Your laptop was running 2.79, so you run the risk of using newer features in the .blend file that your render cluster running older 2.78 cannot cope with.

[@NovaspiritTech]

I tried compiling from the pi thinking it was going to take maybe 1 day... It took closer to 4 days with fixing compiling issues.. and in the end blender didn't want to start

[@lawrencedoliveiro9104]

@@NovaspiritTech Did you follow the building guide wiki.blender.org/wiki/Building_Blender/Linux ? Tip: I use “apt-get build-dep blender” to fetch an initial set of needed dependencies.

[@peterxyz3541]

Thanks for the info! I’m trying to learn blender

[@arthurdent8091]

Hi Don, I like what you do, but Chr*st you blow me away! was trained with big iron, and you seem to have a handle on everything you come across. Damn. I'm really impressed.

[@Highness]

I wonder if the cluster could render OBS Livestream

[@marvinfadri3479]

Hello, do you connect the pi through ethernet or wifi? To use network render?

[@arturomalapitiii3783]

i have a question sir, did you do any configurations to the 4 raspi prior to this? or just connected it to the network and used blender? i have 4 pi3b+ and 4pi4 and is interested in using it as a rendering farm like this.

[@ashrude1071]

Sooo anything with blender can be used as a rendering node? Thank you so much!!

[@GianlucaBrizi]

Hi, I am a subscribed in your channel. I have some SBC like you. You made a collaborative parallel computing cluster between them. I wuld realize it too. Have you published the tutorial somewhere? There are only tutorial to create containers for webservers and clouds. Nothing around parallel rendering. I would like to have the help of SBC connected to the network but I would like to have the monitor attached to only the first, the master.

[@fabriziodutto7508]

It is not possible to use the GPU inside the Raspberry to render, instead of the CPU?

[@chrismofer]

tysm comparing it to your local laptop CPU, and actually amazing results that 4 armv7's beat a quad core i7. what the heck!

[@ultraviolet.catastrophe]

Correct me if I'm wrong, but one Raspberry Pi 3 has 4 cores. The cluster in the video was 4 nodes, giving us a total of 16 cores. However, your analysis is still awesome because a quick Google search shows that the price of a Core i7 processor is $350. On the other hand, a Raspberry Pi 3 goes for $35. Four of them will cost $140. That's less than half the price of the Core i7! Thus, if you have $350, it looks like it's better to just buy ten Raspberry Pi boards giving you a total of 40 cores! Wow. 40 cores would beat an i7 SO BAD!

[@TheBowersj]

Can you network encode videos to h264 like this, some of my vids would take a weekend long to convert.

[@NovaspiritTech]

wow i just found out you can... github.com/ccremer/clustercode

[@TheBowersj]

@@NovaspiritTech wow this is amazing, I finally see why these clusters are being developed by the community. Thank you so much

[@LANgeneral77]

Thank you so much for this I have a laptop with a beefy i7 and ill use this to render while i'm working in blender :)

[@TaiKibris]

Good content. Would you need a higher spec for the master? Or could all be the same

[@NovaspiritTech]

you can run the master on a Pi if you wanted to. but it will require 1 core free from the rpi so in the slave settings you will need to set to use 3 core leaving 1 core free for master

[@TheNewton]

Could this be extended to cheap rooted smartphones as well, say using wifi or usb networking? what's the graphics hardware on Pi's vs an avg $20 smartphone

[@AlexeyZubkov]

I don't care about rendering. Any suggestions how to build transcoding farm with ffmpeg? Hadoop , Kafka? p.s. thanks for your endeavor!

[@hyrumlowry6129]

if you adjusted a few more settings in blender you could have made the animation render quicker on both you pi's and computer

[@eLJaybud]

"I didn't do this years ago because there wasn't an easy tutorial as I don't really know what I'm doing"

[@alexradu1921]

lol.. interesting thought

[@ctrlaltdel1758]

Alex Radu it’s better for hosting docker containers

[@AcielGamingTech]

Hey Don, Which Desktop Environment is that on your i7 Laptop in this video & which OS also?

[@NovaspiritTech]

HAHAHAH i'm glad you notice it looks different!! i'm working on a video about that but since you asked it's running solus & budgie de

[@daanwijns6568]

would like to know as well ^^

[@prussian7]

I'd still like to see the super computer. I would like to use for Monty Carlo Optimizations. Wondering if I can use FORTRAN or would have to use a different language that can be split between CPUs.

[@Akshatgiri]

I wonder if it's possible to create a cluster with android phones. Do you know if it's possible to install raspbian or some other linux distro on andorid devices?

[@nabagaca]

it's possible to install Linux distro's on Android phones, although only certain phones have proper compatibility due to drivers I think (I remember someone ran Minecraft servers on Android phones via a Linux distro), however, phones aren't really made for sustained 100% load and there likely wouldn't be much benefit to it.

[@stale2665]

Indeed. New phones are reasonably fast, but there is one thing that's impossible to get around in that form factor, and that's heat dissipation. This is why phones are fast only for short bursts before they need to cool down. Once the heat has built up inside them, it takes a long time for them to get rid of the heat without any sort of heatsink or active cooling. On single-board computers like Raspberry Pis, there's usually room for adequate cooling, and even fans, if needed. Some phones have some kinds of rudamentary cooling systems, such as a strip of copper (or an extremely thin heat pipe) going from the cpu and across the rest of the phone internally in order to equalize the heat levels across the entire device, but without an external ribbed design like those you see on heatsinks, the dissipation out of the phone is going to be pretty disappointing. Not to mention, most people hold these things in their hands, and it's really uncomfortable to hold a 40 celsius device in your hand for a long time, even if the electronics inside could survive temperatures higher than that.

[@EugeniaLoli]

You don't need to install Linux on Android (or even iOS) to use them as Blender rendering slaves. You just need to port a lesser version of Blender (since itself is open source, only create a gui with the relevant backend code) and have it run as a headless server on a phone. The only problem with this idea is that some Android manufacturers are killing random background services (that Blender would be set as) because they want to free RAM at random times. It's a well known problem with Android phones, and some developers have complained to Google, by asking them to ban these manufacturers from using Google Play store, unless they play nice. No action was taken by Google so far about it though.

[@tomvh]

Very cool project! I wanna try this out for fun. Random question: what theme are you using for chrome?

[@HelamanGile]

Can you cluster smartphones and use them for rendering?

[@Hellhamster]

Thanks for all the clear information, i am a total tech noob and instantly loved raspberry pi but i hate that prompt green.. Can anyone tell me how to change my ...@...$ color to what is used in video?

[@denvera1g1]

I'd like to see an updated version of this using the turingpi and 7 of the pi3+ cpmpiute modules, though i dont think it will be usefull till we see some 7nm 8 core rPi

[@yumri4]

i would like to see which would be better. A Cluster of 7 RPi3+ or the turning pi cluster board full of all 7 compute units of Raspberry pis. The Raspberry pi 3 B+ s will total up to 370.30 USD while the cluster will be 538.93 USD. So going by value the Raspberry pi 3B+ is the better value for the cluster but the Turing pi using 7 Raspberry Pi compute module 3+ 8GB of MMCe onboard flash for the set up for a cluster and no SD cards to go bad nor any SD cards to flash.

[@Supercon57]

I'm curious as to how fast the network render would have ran with chunks set to 1 but regardless cutting your render time in half is pretty notable

[@retocottini6324]

Nice Idea. I briefly rendered the hole thing in my not so new HP Z420 workstation with a PNY Quadro K5000 on the GPU also in Cycles. Rendertime for absolutly the same: 1min 38sec. In addition, I have to agree with algotrash. At the Moment I optimiert a Storm cloud, for this thing Blender uses over 9.9GB. And this only vor a cloud. So its clear use bettet one of the new Nvidia RTX cards, but you see, the bigger the better. The good reason for the not released Tiatan RTX if you do some more complex like landscape or somthing with highpolymesh.

[@marcofranchimoretti69]

Thankfulness Novaspirit for share your experience and idea. Allow me ask you this: where you downloaded the original blender file ( where the render Planet Collision ) ? I will follow your example for processor experiment on the render speed. Thanks and congratulations for the video.

[@iflnr978]

@Novaspirit Tech I would be interested in video of XU4 or best price to power ratio blender and Logic Pro x rendering farm vs latest i7 or 12 core Xeon E5 (MAC Pro). Although this sounds expensive so I don’t know how to fund it. Thanks for the interesting video

[@alissondamasceno2010]

OMG Blender works on a rpi! O.O That's good to know =)

[@nlcirque]

Does this generate any kind of performance metrics for each node? If so this would be a pretty good real world method for benchmarking SBCs.

[@NovaspiritTech]

that's a pretty good idea..... i'm going to look into this...

[@ahmad.ridwan]

I've rpi cluster with 5 node. Im very interesting with ur project. Is there a tutorial written on the blog or web that i can read? Sorry for my english

[@ahusyn]

What's the operating system you are using on the laptop. If its Linux, which version and what theme

[@kitalthevali]

I might try this my tower won't render things and my laptop takes a week to render 600 frames (10s @ 60fps) heaven knows how long 1440 frames would take (10s @ 144fps)

[@gucu]

Thank you for the video. I am looking at building a media server and wonder what sbc you would recommend please?

[@NovaspiritTech]

like a plex? with transcoding?

[@gucu]

@@NovaspiritTech Plex and I do use Handbrake to transcode. Is it possible to hook up a dvd player then?

[@benricok]

@Richard Brown You can use an Udoo X86, it does have th following: 1)sata port 2)good cpu for a sbc 3)m.2 slot 4)gigbit ethernet 5)enough ram(1-8gb)

[@handsomepixel5193]

Based on your comment, you seem to pre-convert your media to a more readable/compatible format for your devices. But I have a question, do you want to do the Handbrake transcoding process on your dedicated PC, or would you rather have your SBC do all of the work for media converting/renaming/serving?

[@gucu]

@@handsomepixel5193 I would rather have it doing all the work. I use my computer for work and whilst it can keep going in the background, it would be nice to let something else do the work and free the work pc. So if possible let the SBC do all the work. Thanks

[@jeffrey8629]

An option for a more "usefull" cluster setup would maybe be something with a docker swarm setup?!?

[@NovaspiritTech]

🤔

[@tberry7348]

Is it possible to run an os on a cluster like that utilizing the entire clusters resources as a single pc? IE can you run Linux or win7 or win 2000 advanced server on the cluster as if it's a single pc with 4* the cores, ram, and gpu power? The next logical step would be to add a hard drive set to each running raid 01. (Hd set = 1tb disk, and 100gb ssd as cache) giving 2tb of backed up storage with a really fast write speed.... Basically I would really like to know how to build a clustered SAS using a low power consumption sbc. The gigabit Port of a pi is more like a 250mbs port but with enough of them in this configuration you should be able to saturate a gigabit switch....

[@pppianissimo]

This won't work for multiple reasons: 1. Cache Coherence You can't just stack up ARM-Processors to get more cores. They use a technology called "Cache Coherence" which essentially means that even though each core has their own cache it will automatically update the cache upon memory access to avoid differences in the data called. I won't get into detail, but to do this, an internal mechanism within the processor has to set a flag that the memory has been changed at that address. This feature is global, which means that it works within all cores at the same time. If you were to split up the processors in a node, you'd need a procedure to again globalize this. Some processors can do this with specialized motherboards (like Intel Xeon), but there would be no feasible solution on a Pi cluster. This would mean that either the caches of all processors differed greatly, which leads to calculation errors, or that you needed to turn off all caches completely, which would make the processors run extremely slow, being bottlenecked by the memory. This leads us nicely to the second problem: 2. Memory speed It is theoretically possible to stack up memory of different Pis and remap memory access cleverly to get global memory readings. There's only one small problem: The speed. I don't know how big the bandwidth of the Pis RAM exactly is, but I'd guess it is able to write about 5 GB/s. The Ethernet of the Pi 3+ is 1Gbit/s, which means only 125MB/s transfer speed between the devices at best - 40x slower than the RAM. Sounds bad? Well, it gets worse. There is a reason we call it "Random access memory" - RAM is designed to take millions of random single calls per second. It can do that because the latency between processor and memory is extremely short, to be precise: 1/900,000,000s in case of the Pi 3+, since the RAM runs at 900MHz. Unfortunately, Ethernet and USB work very differently. They send data packed into large "packages" and don't wait for a response. This is because they have a huge latency compared to the RAM/CPU-connection. You won't get any connection via LAN or USB that has a smaller latency than 5 ms between two processors. This doesn't sound like much, but it would lead to your cluster RAM essentially running at no more than 200Hz - much slower than even a 1970s PC. No program with even a small number of random accesses would run even close to normal speed. This is by the way the reason why Nvidia SLI and AMD crossfire aren't able to share the VRAM of two gpus. The same problem would obviously occur with the GPUs inside your cluster. 3. No OS There isn't any operating system out there that could run on such a cluster. All known OSs are designed to run either on a single CPU board (or a multi CPU-board, but in that case the specialized hardware will just PRETEND to have a single processor) or on a Super-computer with hardware very different to a RPi-Cluster. It's not impossible, though, to program such an OS, but you'd have to start from scratch since you couldn't even use the Linux kernel. I could name even more problems with your idea, but I think those few already show that it's impossible. That doesn't mean, though, that you can't use an RPi-Cluster for supercomputing tasks. But certainly not in the way you imagine it.

[@ZonaALG]

I always wondered if a rpi cluster would ever work for something... Comparing expensive machines with expensive rpi cluster pc isnt probably the best idea. You should compare a laptop/simple pc with 5rpi cluster instead, what would be the cheapest-most powerful configuration for clustering rpis?

[@RedMageGaming]

Every "Cluster" is essentially configured the same way, its the software that really changes things. That's why there are so many "how to build a Pi Cluster" videos out there that all just hook 5 or so Pis up to a network switch/router and then go tada its done. For the comparison to mean anything a task must be selected. Weather it be something as straight forward as this mini render farm, or a cluster of Pi's running Tesserect for processing scanned in documents, to something like a web server cluster, distributing functions such as SQL Database, File Storage, web application hosting, and load balancing etc. But if we for arguments sake stick with a 5pi cluster, we essentially have 20 cores to do work, if an outside machine is directing the action, like in this example his laptop sent the work to the pis, typical stand alone pi cluster though, one of the pi's is the master. so 4 pi's doing work, one pi directing the show, making this both as cost effective as possible and as powerful as can be, your looking at your 5 pi's a 6 port network switch, or a wireless router if you want to wirelessly network the pis, and a USB power supply capable of supplying about 2 amps to each pi at a stable 5 volts. [never cheap out on your power supplies.] I personally recommend going for a wired network switch for stability. Cables Cables Cables. with everything together for a 5 pi cluster your looking at around $200 to $250 USD depending on what bells and whistles you add to the construction, weather you went with a nice little acrylic stacking pi case, or just put inch long nylon standoffs between the pi's and weather or not you put fans on it and whatnot. At that price point, the only computers your really able to compare to at the same price point are mainly netbooks/ultrabooks from acer and HP that are all invariably running atom or celerons that barely run windows 10 on 32GB EMMC drives. But I think that actually brings a better image into mind about exactly what we're comparing here. I think any distributable task that doesnt need a 3 to 4 Ghz proc to perform, will just complete faster on a cluster of pi's than a single desktop/laptop at the same or id almost say twice the price of the pi clusters base parts. And in many cases, probably use less power as well.

[@ZonaALG]

@@RedMageGaming but what if for example we compare that cluster to a core i3 3ghz (2 cores 4 threads) 4gb ddr4/ddr3 laptop? (which is a kind of a standard config on low budget used laptops)

[@RedMageGaming]

Then again it comes down to usage, for singular tasks that benefit more on single core performance or needs more than 1gb of ram to even perform a task in general than that laptop would be more reasonable. However for tasks that greatly benefit from parralel processing over single thread performance the pi cluster would greatly out perform the laptop. Like this rendering example, while it is the most basic use of rendering and super complicated scenes likely wouldnt go over as well if the resources to render were more than a gig. But the 16 core cluster here performed as expexted compared with that laptop, and the laptop hes using iirc isnt a terribly cheap one either.

[@truecommentor2657]

Why you save slave.blend in each slave and you started rendering without any command to slaves can you explain me please

[@lodesmets9815]

I don’t know if you do any programming, but is it also possible to compile on a cluster?

[@NovaspiritTech]

yes i do programming, and yes you can cluster compile with gcc

[@NovaspiritTech]

checkout icecream - github.com/icecc/icecream

[@hypersonicmonkeybrains3418]

im banking on a Zen2 16 core well over 4.0GHZ to come out next year.

[@hypersonicmonkeybrains3418]

Then again, maybe AMD will have some amazing priced 8GB video cards and i could get 2 of them.

[@rkg456]

what model Pi's were you using in the render farm?

[@pcmaster888]

I never got the Network Rendering Plugin to work in Blender 2.79.. I always thought it's not getting developed further anymore and replaced soon

[@CalebDaThing]

Can you use a 2 or 3 node raspberry pi for gaming?

[@naizeylines]

Its a cheap alternative if you need your main pc while the farm is rendering

[@faqtao]

Hey ! Tell me how made rendering simply video? Sony Vegas rendering ?

[@adityakadali]

6:12 Your computer facing screen tearing issue. (Breaking of frames)

[@NovaspiritTech]

yea obs was messing up at those times....

[@techwizpc4484]

I also experience those tears. I suspect it's the network speed also being a problem and I'm limited to 100Mbps on my LAN connections.

[@adityakadali]

@@techwizpc4484 It's due to graphic driver issue.

[@YoloMonstaaa]

I think it's a Linux issue.

[@misaalanshori]

Could be slower because you're recording it, but the more cores probably helps too

[@TheNewton]

Wonder if bmd's fusion could be made to network render this way as well since davinci resolves popularity is taking off.

[@primemover84]

Which Linux distro are you using? I really like how it looks!

[@NovaspiritTech]

Ubuntu budgie

[@fasilahamed6306]

What os you're using this video

[@bearwolffish]

very cool

[@justethom]

I didn't know that we could run blender on Raspberry Pi !

[@lawrencedoliveiro9104]

I’m sure it will compile on any architecture that GCC can generate code for.

[@MrCcfly]

wasnt cycles opencl too ? so u can use gpu inside v8 arm or not ?

[@kingmasterlord]

you know those cell towers disguised to look like trees? wifi trees. solar berry bushes where you stick these on those solar panels that fold out like flowers.