I9-9900K upgrade

[HenrikGrosen]

Hi,

I am looking for insight for improving Metashape performance.
I have looked at the various information at:
https://www.pugetsystems.com/recommended/Recommended-Systems-for-Agisoft-Metashape-145/Hardware-Recommendations

Our current setup is: I9-9900K @ 3.6GHz (8 cores, 16 threads), 64GB ram, RTX 2070 super
Currently I am running a job with 60.000 24MP images, and it is have a hard time coping with that.

From the looks of the benchmarks on pugetsystems it does not appear that upgrading to a high core count cpu will help (ex. AMD Ryzen Threadripper 2970WX @ 3 Ghz, 24 cores/48 threads).
And the I9 10900K is only a minor upgrade to our current CPU.

What can be done?
Is a cluster setup the only way forward, and what performance increase can be expected with such a setup?

Looking forward to hearing your input/experience.

Best regards,
Henrik

[Arie]

Hi Henrik.
I've just been in a similiar position on upgrading my PC and I had the same questions regarding the benefits of high core count.
In the end, I went with the Threadripper 3960X (24 core). On the one hand, it is very convenient to be able to just assign half of the cores to Metashape and being able to continue working on something else without much lag (I have a dedicated GPU just for display).

If you work with chunks, you can process multiple chunks in separate instances of Metashape, which increases perfomance (somewhere at Pudget that possibility was even mentioned).  In addition, I work with other software (Blender, GIS etc.) which can utilize the cores more efficiently. And I would expect Agisoft et al. to optimize their algorithms for larger amount of cores in the near future since high core count has just recently become so widespread. Also, the amount of PCIe lanes the TRX40 platforms offers can be quite useful when working with many M.2/ PCIe SSDs.

If you have any questions, feel free to ask away.
Cheers!

p.s. for a cluster you will need a separate license for each node.

[Alexey Pasumansky]

Hello Henrik,

I think you should check more detailed timing information corresponding to each processing stage:
https://www.pugetsystems.com/labs/articles/Agisoft-Metashape-CPU-Performance-AMD-Threadripper-3990X-64-Core-1665/

I am not quite sure, but I'd rather assume that the mesh and tiled model generation in those tests used dense cloud source option, which took longer time on AMD processors.

In the version 1.6 we usually recommend to use depth maps based mesh generation option. Also keep in mind that image matching and depth maps generation are GPU supported and powerful GPUs are tens times faster that even faster GPU, so it is reasonable to pay attention to the CPU-only steps.

[HenrikGrosen]

Thank you for your comments Arie and Alexey,

Arie, I can see why you went with the 3960X, it seems like the sweet spot for getting a high core count while ensuring a high clock speed, which out pricing going through the roof.
However, I think I will wait with our CPU upgrade, as it seems to not give a big upgrade on our 9900K, unless you run multiple Agisoft instances.

I am leaning towards adding another RTX 2070 super instead. Alexey, is NVlink required between the two graphics cards, or do you just need them standalone/side-by-side?

Best regards,
Henrik

[Alexey Pasumansky]

Hello Henrik,

You do not need to connect the graphic cards, thus you can actually use different GPU models in the same case, as no SLI/Cross-fire connection is required.

[DriesS]

I'm also thinking that 64GB of RAM might be serious chokepoint when you're working with ±360GB of pictures, no? 

[jmos]

Hi, recently upgraded to the following and have been quite happy with the performance.

AMD Ryzen Threadripper 3970X 32 Core 64 Thread 128MB sTRX4 Processor
OS Drive Samsung 970 Pro Series 1TB M.2 NVMe SSD
G.Skill 32GB (2x16GB) DDR4 Trident Z RGB Neo C16 3200Mhz
2 x GeForce RTX2080 Super XC Gaming 8GB GDDR6

John