I am trying to offload rendering work from my local machine running UE4 Swarm[0] onto one or many AWS instances (note that Swarm only utilizes CPU and cannot use GPUs to process lightmass data yet).

Right now I am poking around testing performance of the different cpu/memory optimized instances. My biggest question I am trying to answer is if it makes more sense to have 128 t2 instances with 1 vCPU or just one x1.32xlarge instance with 128 vCPUs. My hunch is that it is better to just have one instance to reduce the complexity of having to manage 128 separate instances. As well, I bet 1 x1.32xlarge instance effectively encapsulates what 128 t2 micro instances would. So far runs of a single instance of Swarm can effectively distribute work across the many vCPUs on the single instance.

If the x1.32xlarge is $13/hour then my team is looking at a $10,000 a month bill from AWS (with the assumption that we must leave it running 24/7 because we are a remote company with teammates across the globe and it needs to be instantly on demand). I feel we could probably roll our own solution at this price with just the budget for 1 month of a x1.32xlarge instance. But I'm not sure what hardware is backing such an instance and what it would cost for us to build our own?

My bet is we could probably purchase sixteen 8 core Xeons and keep it under the 1 month cost of renting the x1 AWS instance.

I found a build on PC part picker that I think takes advantage of the low cost of old Xeon processors:

https://pcpartpicker.com/guide/dfkcCJ/1500-dual-xeon-rendering-pc

So maybe we just spring for 8 of these but this kind of form factor is not ideal to scale so hoping someone might suggest a server box/rack.

Am I on the right track here or would you guys recommend an entirely different approach to solving this problem?

Thank you!

[0] - https://docs.unrealengine.com/latest/INT/Engine/Rendering/LightingAndShadows/Lightmass/