DIY Renderfarm

Hello Lainons. I've been a 3D artist/CAD engineer for the last five years and only recently decided to give up on all the cloud render services and make my own render farm. I understand easiest and cheapest way to do it, is to build Motherboard-CPU-GPU-Storage-PSU combos and link them up. Now consider this situation: Around five years ago I read an article where around 12 Samsung devices were used to create an experimental mining rig for Bitcoin. My question is, can I use a similar setup to create a large renderfarm made up of mobile devices. Also, if it is possible, how do I go about creating such a setup? Other possible ways to create a render farm are also welcome.

Buy a GTXRTX4099TISUPER or whatever the fuarrrk its going to be called when it comes out, put it into a pc, and youre basically done :)

>>8366 instructions unclear, chinese 350W PSU made the gpu explode

don't just use the gpus that are good for bitcoin/crypto, they're not always the same as the good ones for rendering. research your own rendering engine and try and find the better gpus that exist for it specifically.

>>8366 Kek bro I would but unfortunately having a greater number of nodes that can render different frames simultaneously is better than having one super powered node that renders one frame at once. >>8368 I know what you mean, I simply meant that if the cpus were strong enough to calculate bitcoin hashes maybe they are strong enough to render pixels as well.

What kind of rendering you do? There are different requirement for different types. Precise hardware is slower than error prone gamer grade. If you need real time you need fast hardware tho. Building a rig is mostly about balancing time and precision according to your needs and budget. You cannot expect precision of NVIDIA Quadro cards from gamaer grade NVIDIA cards. You cannot expect speed of gamer grade cards from Quadro cards. You cannot expect Intel Xenon precision from desktop grade Intel CPUs. You cannot expect desktop grade Intel CPU speed from Intel Xenons. You can split some rendering to CPUs and some to GPUs. (In blender for example) You are mostly unlikely to render part of frames of same scene split on CPUs and GPUs without getting different calculation results for movement/physics in that scene depending on what it was rendered. That might result in impossibility to merge frames together. etc. Also if you are gong to do a really looooong renders you might be interested in server grade ECC corrected memory. So if you want an idiotproof rig - keep hardware as unified as you can and render whole scenes on same type of hardware. If you need to render looong and precise (pretty much everyone who is not a hobbyst) go for server/workstation type of hardware. If you can live with imperfections and not exactly same results of rendering same scene on same hardware - you might take a risk with gamer grade stuff - but remember you been warned.

>>8374 The fuarrrk do you mean it's error prone?

>>8378 He means ECC ram. If a Ram bit flips in your game, maybe a pixel flickers. Nobody notices. But if you iterate over 20GB of model data in ram, a flipped bit can have a cascade effect.

>>8378 computer hardware like CPU, RAM etc. have it's limits in amount of time they work correctly. Once in a while CPU fails to operate correctly. There are various possible reasons. Some are internal properties of materials used that start to fail under certain conditions - it might be voltage, it might be frequency of clock signal or reaching certain temperatures. Other are external - like outer space radiation, solar radiation bursts etc. If it's fail when you are playing Roblox or Fortnite worst you can get is BSOD or a freeze. Not a big deal. If it fails when you are calculating weather forecast for the military - well it might have dire consequences. That's why people doing weather forecasts prefer to get more reliable hardware. That's why they use lower clock speed that gamers would find pathetic in their CPUs - there is (or at least was for quite a while) a relation between cpu clock frequency and reliable work. As gamers do not typically depend as heavily on precision of calculations and an error or two once per few hundred frames is fine for them they do not need CPU with same level of error proof working. There are also various error correction mechanisms that might be implemented in hardware dealing with data - ECC is one of them typically used in RAM chips. RAM chips are notorious victims of bit flip errors. There are many stories about bit flips in history of computing having it's consequences in real life. ECC aims at making bit flips much much much less likely to happen.

>>8374 Yes I understand that different forms of rendering and different forms of tasks that you can utilize a CPU and a GPU or maybe even a TPU vary a lot. I'm not calculating bitcoin hashes here or trying to train an AI using a dataset. This is specifically for general 3D render production and animation using simple physical based render engines like Redshift and Arnold and Cycles. You're absolutely right about what you're saying, having a single system render a frame would produce more coherent results for multiple frames. The only utility here I'm trying to achieve with a render farm is the simultaneous number of frames that I can render at a specific time. Higher are the number of nodes on which I'm rendering, higher will be the efficiency at which I can deliver. From what I have observed from experimenting around is that having multiple similar nodes to produce multiple frames will produce more coherent results. Higher is the memory of the hardware you're using for render, better is the output. Which is why renders produced via CPU will always have a higher quality over renders produced with a GPU. A threadripper CPU will be very efficient in render production whereas an Intel Xeon will suck bootyhole. My end game here is to create a renderfarm with a fairly good efficiency and not be very expensive.

>>8410 You seem to understand topic fairly well. Please specify what is your exact area of doubts/questions. This is unclear/ambiguous. Are you unsure about some types of components or are looking for any specific price/precision/speed ratio? Or you are looking for tips on how to organize units in a network. Or something else.

>>8413 Yes Sir, all of the things you just mentioned. I have never actually done such a thing myself and since this will be a fairly expensive task even though I plan to use as cost effective hardware as possible, I keep wondering things such as render speeds achievable by CPUs and GPUs of previous generation. Even if it takes a bit of extra time, how good or bad will it be if I used a CPU like say Athlon 3000g or fx 6300 or even say core 2 duo, or if I used a GPU which only has a vram of 2gigs or 4gigs but is a almost a decade old ? If I make such a build, is there a way that I can link all of these different nodes together ? How do I stack them ? Should I get separate cases for each of these nodes or should I use a single very large rack mounts and stuff ? What sort of heat management systems will I require ? I'm sorry if these are noob questions I just don't know where else to ask for help on these sort of subjects.

>>8414 Intel been adding a lot of floating point vector operations instruction sets to their CPUs with every generation. AMD been following. Keep in mind that older CPUs might be cheaper but are slower and can consume more electricity to do the same job. Especially if they are made in older "bigger in nanometers" technologies. So you might pay less for hardware but due to slower render times and bigger power consumption you can spend few times more on electricity bills. If you are planing to use your render farm for few years to come it might be not worth to look into such ancient hardware because of that. As for GPUs GPU are architecture designed for algebraic operations on vectors that GPU needs to do a lot in order to calculate all that stuff needed for rendering. How to pair all of that I'm afraid I cannot help you in a way other than some general opinions. I do not keep up with latest market products. Think about what is your realistic budget you can spend on your farm - then try to find some benchmarks of components that are of your interest - best if they are done on software you are working with and think about how you are going to organize rendering. Then do a lot of math calculating consumption of power, speed per node and component cost per node proportional to speed and power consumption. You potentially might need to understand how your software works on multi node environment. I mean if every node receives at the start of rendering info about whole set of scenes and is just doing rendering of individual frames and keeps rendered frames on own disc and does not need to transfer huge amount of data fast and often you do not need crazy fast ethernet cards. Think "what is done, how often, how big amount of data is typically transfered between nodes". Don't expect yourself to get everything right. Even pros can be surprised about their creations. Sometimes positively, sometimes negatively.

>>8420 Understood. I'll keep these things in mind thank you. I suppose it does make a lot of sense because the speed at which the graphics industry is moving, my node setup would be useless in the next two years. >>8368 Also I tried experimenting with blender on a mobile phone and it worked beautifully. Now I know this isn't a practically viable option but the fact that I could do it was pretty epic.

>>8429 hey OP, how's your farm going?

>>9463 Pretty well actually thanks! I went with 4 nodes each with 2x Xeon 2670 v3. The GPUs are different thought, two of them are Titan X Pascal, one has a 1070Ti, and one has a 1660 Super. I just decided to go with used parts from eBay for the first go at this whole renderfarm thing. Now that I have a much better understanding of it though, I'm sure the next farm will be much better haha

Also to whomsoever it may concern, I have come to understand that a larger number of cores is a much better rendering hardware. So if you want to choose a CPU for 3D/CAD work, its better to find a CPU with a larger number of cores. You can find Xeons and Threadrippers and sometimes even EPYC CPUs on ebay discarded by servers that are fully functional and cheap. You don't have to spend a huge amount of money in order to make it work by yourself. As for GPU, the prices, my friends, are very low at this point in time. Go to eBay and look for your GPUs. You can always find benchmarks for GPUs and CPUs at https://opendata.blender.org/benchmarks/query/ and whatever is good for blender is almost always good for other graphically intensive applications involving CAD. Remember that realtime graphical performance is not the same thing so stop looking at video game benchmarks on youtube they will not tell you what is good for you.