I have a render farm here in our office that consists of 4 x dual xeon nodes with 8 x Geforce 1080ti cards on each node. They currently have 64GB ram but that can be scaled massively.
I am looking at very large city scale datasets and want to know if it is possible to utilise at least one of these nodes with all 8 x graphics cards to speed up processing and is there a standalone processing app (similar to Octane Render for example) that is in progress to assist with the calculations.
First, here you can read more about our minimal HW requirements.
Generally, most of the processes in RealityCapture are out-of-core apart (not dependent on the RAM) from the Alignment process (registration of cameras).
RealityCapture will use all available RAM if it leads to faster computation. Otherwise, it splits jobs so that it fills into the Windows System memory (SWAP). So technically 16GB is enough for reconstruction, texturing, etc - but more RAM could lead to faster processing.
Memory consumption during the alignment phase depends on the number of images (not size) and the number of detected features per image. For the default setting of 40 000 features per image (Alignment setting), you can expect the following boundaries:
2,000 images - 16GB RAM
4,000 images - 32GB RAM
8,000 images - 64GB RAM
16,000 images - 128GB RAM
By decreasing the number of detected features to half you can approximately decrease the memory consumption by half as well. The approximate formula is: RAM = features x images x 200 bytes.
Thank you for explaining the MEMORY consumption and boundaries for alignment. That helps a bit.
Is pushing tasks to CALCULATE alignment in parallel on the GPUs not a good idea?
GPU is used during Depth Maps calculations, after this process is done it RC will calculate meshes and this process is done by CPU. So, GPU is not used for alignment.
