Heavy scenes "bottle necks" (CPU/RAM)

[yaa]

Hi everyone

Our group is need to build large heavy scenes (50x50 km and more) with a detailed real landscape, dense urban development, vegetation, road network, a large number of dynamic objects, etc.
There are two rigs for this:
1. Intel i5 8400 (6 cores) + 64 GB DDR4
2. Intel Xeon e5-2643 (4 cores / 8HT) + 16 GB DDR3 on Dell 0TF3RV server motherboard (http://www.dell.com/support/article/ru/ru/rubsdc/sln290764/the-hardware-specifications-for-the-precision-t7600-desktop-workstation?lang=en)

In the second rig, we can add a second Xeon processor and increase the memory up to 128GB
The first rig can not be upgraded.

Question. On which rig would it be better to build large heavy scenes? What will be the "bottle neck" when building a scene: processor frequency, number of cores or RAM?

We need to understand: need we are upgrade the 2nd computer and build scenes on it or buy a new computer similar to the first rig?

[Ihde.Sascha_]

What is your GPU?

[yaa]

9 minutes ago, Ihde.Sascha said:

What is your GPU?

1. GTX1050Ti

2. GTX680 (upgradable)

[silent]

Yaa,

It really depends on scene organization. Nodes baked into the MeshCluster / WorldCluster can give a huge performance boost on GPU (but loading CPU more). Usage of occulders inside buildings can also give a GPU boost, usage of LODs and other techniques explained here: https://developer.unigine.com/en/docs/2.6.1/content/optimization/ can also give you more fps without changing in hardware.

You need to check with engine profiler (show_profiler 2) where is your bottleneck. In 2.7 update we will extend microprofile support to enable GPU timings reading, so you will be able to see which exact rendering feature costs on CPU and GPU.

Adding another CPU would not help, it's better to scale the GPU first (at least to the some modern GTX 1080+) and then see the results. We are constantly improving the CPU usage and some of the objects are already internally support different threads, but mostly we are utilizing a single CPU core.

Thanks!