Render farm with multiple nodes vs. one server with multiple cheap GPUs?

Hi everyone! I’m trying to plan out a future build to use for rendering on something other than my production machine, since I suspect that costs for using cloud render services will stack up quite a lot over time.

The two things I am considering are:
1) Buying a Helmer drawer cabinet from IKEA and building a number of render nodes to use with the Crowdrender plugin, each with at least 16GB of RAM I’d guess, and with a decent-but-not-too-expensive CPU – I am currently looking at the Ryzen 1600AF or 3300X or similar, and then running them headlessly (without GPU or GUI) only for rendering.
OR
2) Buying a motherboard with a ton of PCIe 1x slots and stacking it full of as many relatively low budget and low-TDP GPUs as I can get my grubby lil’ mitts on, crypto mining style. Since Blender plays nicer with CUDA than OpenCL I guess I’d go for something in the GTX 1050 Ti class or similar.

Both of these options have the upside that I can start small and expand them slowly. The CPU solution also has the upside that I can run tasks on a given node that aren’t run on the GPU, which is cool.
I noticed on some benchmarks (CPU vs. GPU) that they seem to do well in different tasks, specifically the CPU seems to do better in scenes like koro, fishy_cat, and pavillon_barcelona, while the GPU does better in bmw27 and classroom.

So, to the question: is either alternative a reasonable solution? Am I overlooking something completely, like the limited memory of the GPUs being an enormous issue I’m not quite getting my head around? Any advice is greatly appreciated, I’ve tried to find as much information as possible both on the forum and elsewhere and a lot of it seems outdated and a lot of advice given seems contradictory.

Thanks in advance! :slight_smile:

Hi, I would recomend one server with a lot of GPUs, because then you have to buy just one Motherboard and one “RAM” …

I’ve done it nearly the same way and have build my own serve with around 100 cores, 500 GB RAM and 20.000 CUDA-cores (=> one big maschine + 2 smal ones))

For example you go with the 3960X, 128 GB RAM and 4x RTX 2060
(enough CPU-power for physics-stuff (24 cores) but a lot of GPU-power (8.000 CUDA-cores) for typical rendering, too)

I hope it helped a bit

Have a great day,
Nico Weiberg