My current desktop has a EVGA RTX 2060 Super, Ryzen 7 2700x & 32GB of DDR4 3200 RAM. For viewport rendering I use Eevee since it’s quicker & has real-time ray tracing, but for final rendering (sending out the completed animation) will I see a huge decrease in render time if I build a render node on my home network of 4 RX580’s utilizing OpenCL? (All are the exact same card.) I will be using Cycles for final renderings for accuracy & quality.
I will be creating 10-30 second 2d/3d clips now, but later on I will be exporting 2D/3D short films (5-30 min shorts) & VFX at 24FPS and sending it to the render node. Just trying to see if my final render time will be improved by at least 50% if I build a separate render node of 4 RX580’s that’s only purpose is to be sent the work from my desktop & render final animations.
The 2D scenes won’t be extremely complicated and the animation style & length of 11 minutes could be compared to an episode of Adventure Time or Regular Show or a 5 minute Studio Ghibli-like short. (e.g. https://youtu.be/xu_TVsBRsOA) The 3D scenes will be more complex but the animation style & length will be similar to a 10 minute Pixar short film, (e.g. https://youtu.be/dOsxkzXoN2A), which I see takes an extremely long time to render.
Just in terms of pure rendering power & rendering each image as quick/accurately as possible would this render node at least cut my render time in half? A monitor will be used once for setting up Ubuntu & then the node will be headless so it won’t interfere with the GPU rendering. I haven’t ran into any VRAM issues yet and each of the cards has 8GB. CPU will be a i3-9100F to support the 4 PCIe 3.0 x 16 cards, storage will be a 500 GB 2.5” SATA III SSD or 500 GB nVME SSD, & RAM will be 8GB (1x8GB) DDR4 2400Mhz. Would having at least 16GB of RAM speed up the process since I’m delivering it from my main desktop over the local network or is VRAM & amount of GPUs more important in this case? Thanks!
Awesome! Thank you for linking those stats. Since I’m rendering animations/VFX & not only still images, is there a way to and would it speed up the process if I had all 4 GPU’s rendering the same frame at the same time or all 4 GPU’s rendering different frames at the same time? Also, my reason for getting a low powered CPU & 1 stick of 8GB ram is that they’re just to get the system up & all the real work will be happening on the GPU’s. I’ll only install Ubuntu & Blender on the node. I see in the article it says:
“You don’t want to go too “low-end” when choosing a CPU, even if you primarily render to the GPU in Blender.”
But since I’m using 4 GPU’s, would that still apply? And that “things like physics simulations are single core processes, but that will be changed in the 2.90 release.” It also says
“We’d suggest a bare minimum of eight cores nowadays, to give yourself some breathing room. Blender probably isn’t the only thing you do on your workstation, so any CPU choice should also hinge on those other workloads.”
But this system I’m making will be used solely for rendering. Will the choice of CPU or RAM bottleneck me in the future somehow?
Well, how much will this render node cost? I don’t think this is a vfm option. 4 gpus to just half the time of renders compared to 1x2060 Super. Have you taken into account the power consumption?
I’d suggest checking out Blender Open Data benchmark site.
Mean rendering time of 2060 Super is 191 while RX 580 is 271. So 4 x RX580 cards “assuming” a good scaling factor is about 2.8 times faster. I believe though the efficiency drops with the more cards you add so additional research is needed. I’d agree with @birdnamnam that you should consider power and thermals! Another consideration is CPU use climbs as you increase the number of GPU cards. So choose your CPU wisely. See this thread for more info:
Good luck and have fun building your new system and happy Blending!
CPU and RAM during render are purely used to store and BVM calculations. As such though they can help, in the end even a old r5 1600 and 8GB of memory would be sufficient to push the GPU’s to render as needed. (pending how large the scene your are rendering that is.
As others ahve mentioned, if you are planning to purchase the GPU’s, 4 rx 580’s I would reconsider with more RTX 2060 super cards.
Else if you already own the cards and just need a CPU+Memory to build a render node, then as metnioned above you’ll be fine with basic system.
2070 Super on your system will give almost the same speed as 4x RX580 with much less power consumption and if you are not buying 2nd hand it will even be cheaper (I’m not talking about even the cost of power supply and the rest of the rig)
If you want insane speed get 2x 2070 super it will probably cost the same as a rendering node.
The specs I take note of lighting, Mrender & Gravity, Reflection & Parallax. I’ve seen posts around the Internet say that AMD is better when scaling, but which specs make this possible?
I know having an entire scene fit into VRAM is one of the most important factors in rendering with complex scenes so the GPU won’t crash and I’m starting to learn more about the all the aspects like memory speed & floating point performance while looking into the specific architecture. Are there any specs I’m giving too much attention or that I should pay more attention too in terms of making a pure render node?
And tbh you’re probably right, I think I may have to just wait for the upcoming RTX cards. I also want to see how switching to Windows for the Studio Driver will improve it as well.
Very great insight with that post, thank you! I will probably get a 8 core AMD I can find in that case. For my build I’m considering the “chenbro rm42300-f” & rosewill rsv-L4000C. Cooling-wise, would the first one be better to get? Also, would water-cooling improve help with performance/longevity? I think that might be overkill tho for a render node tho. I’m going to get a cart to house more of the nodes when I can make this one good.
If I end up with blower cards, I want to get this case: Rosewill rsv-L4000C (https://www.amazon.com/Rosewill-Rackmount-Solution-Building-Ethereum/dp/B074W6T178/) and would install with risers.
If I end up with open-air cards, I want to get this case:
Chenbro rm42300-f (https://www.amazon.com/dp/B004GCOY3M) and would install directly onto the motherboard. Would these be good cooling-wise for the types of cards I would end up with? Are there any better options for rackmounts or any other type of chassis cooling-wise? I’m going to get a cart to house more after I build & use this one for a while.
Gotcha. I think I’m going to get a low-cost 8 core AMD CPU. Going to look more into the 3XXX series and see if it’s worth waiting for. Thanks a lot for your insights!
Thanks, I should have considered power consumption more before! I’m looking into the speculated specs of the RTX 3XXX, what’s your opinion on the upcoming release?
Hi @thaizoo, sorry for the slow response I’ve been away camping. I would agree with @anaho that if you can hold off and wait for the new announcements from AMD it might be worth your while. Intel is going to announce something soon also (see https://cogconnected.com/2020/07/intel-something-big-september-2020/). I’d guess something to try to hold back the torrent that is AMD right now. Not sure if it will pan out. But all of this is great for consumers as pricing is most likely only to get better. Competition is great for prices as is a new CPU generation as current CPU gens costs will drop. Those bitcoin server cases are interesting but you might want to double check the reviews and Q&A. Seems the Rosewill might have thermal issues from what I see. The Chenbro seems highly regarded and not just on Amazon. As for water cooling that might get expensive quickly especially if you want several GPU’s + CPU. If you are good at tinkering then water cooling might be fun for you but I’m pretty sure these server cases aren’t designed with water cooling in mind. Happy hunting.
