Hardware optimisation and memory channels

[jrp]

I’m specifying a workstation around Metashape’s capabilities, and have a question.

Puget systems recently tested CPUs, and found that the intel 10900k outperforms the 10900x by a significant margin. (https://www.pugetsystems.com/labs/articles/Agisoft-Metashape-Performance-Intel-Core-10th-Gen-vs-AMD-Ryzen-3rd-Gen-1765/)

This is great as the associated hardware is cheaper because the 10900k uses more mainstream motherboards/chipsets, a different socket and is generally more accessible.

The 10900k has 2 memory channels, the 10900x has 4. I had always assumed that memory channels and memory bandwidth would be such a fundamental requirement for performance in Metatshape that I had never previously considered any of the consumer hardware before. I did a full double take when I saw this data.

My workflow is small object museum work, where we are sometimes having to do 600+ 24MP images per object due to a variety of reasons. A machine with 64GB of ram once ran out of memory on the largest models we’ve done.

My question is: are Puget’s findings general, or are they specific to type of testing they perform? The large project (park map) is 792 images at 18MP (so actually pretty close) – but it’s such a different dataset to ours, I’m not sure it applies directly.

I’m torn between buying a 10900x with 4 memory channels as my intuition says, or buy a 10900k with 2 memory channels like the data suggests.

Have I missed something here?

Thanks,
JRP

[wojtek]

Memory channels won't matter that much, you can get a 2 channel platform and then fit it with 4x32gb of ram for the 128gb total. That should be enough for most depth map based projects (dense cloud part is also getting out of core in 1.7.0 it seems, though)

[jrp]

To be clear, it's not the capacity I'm woried about, it's the bandwidth.

My understanding is that half the channels leads to half the access rate.

Thanks,
JRP