Computer Build Question

[corey.forster]

First, I simply would like to say that this truly is an amazing tool. Thanks for going the GNU route!

I’m primarily interested in using CloudCompare for combining many laser scans together to perform a historical analysis of the overall dataset. My issue is the computer I initially had access to barely was capable of handling a single one of these datasets let alone multiple. Rough numbers are millions of data points per scan, collected roughly hourly, nearly 24/7.

I’ve done some basic searches of the forum and have discovered the following so far...

CPU - I’ve seen mentioned that due to calculations being relatively easy the CPU speed isn’t as important as overall number of cores. Would 8-core 16-thread be sufficient? A coworker pointed me towards the Intel i9-9980XE Skylake X 18-Core although this is quite an expensive route without understanding more of the benefit part of the cost/benefit. Thoughts?

RAM - more the better right? I recall someone processing some satellite lidar was running 128Gb. Seems like perhaps starting with 64Gb, but having space to upgrade to 128Gb would be a good move?

GPU - another key piece of hardware and internal memory is clearly important. A coworker spec’d a GIGABYTE GeForce RTX 2080 8Gb, but also found a ASUS ROG Strix GeForce RTX 2080 with 11Gb. Again thoughts?

My current conundrum is not having a computer capable of the datasets I have access to, but also the unknown of not understanding how much of an impact these various components will make. I do think I easily have the potential to approach the maximum datapoint limit (something like 4 billion, right) and I’d love to not require a ton of statistical sampling.

Any help/guidance would be greatly appreciated and I hope to become a contributor to the community some day!

Kind regards,
CF

[daniel]

Hi,

Well those are probably all good choices, and I would even keep the lower choice each time (which is already very good - better than what I personally have ;) ). Then it depends of course on the actual number of points. But with 8 cores and a GTX 1080 I can already handle clouds of several 100M without too much issues.

If you want to work with 500M to 1G points, then you'll need indeed to increase the RAM (and anyway take some coffee breaks ;)

[CloudyCloud]

Since not all operations are multithreaded, I would guess that i9-9900K or i9-10900K might bring better overall performance than i9-9980XE due to higher single core speed.
When it comes to GPU, if I recall correctly, TI versions have slightly higher performance on top of more VRAM, but I'm not 100% sure. Check Nvidia specifications to confirm it.

[WargodHernandez]

Currently the point cloud limit is defined like this

Code: Select all

//! Max number of points per cloud (point cloud will be chunked above this limit)
#if defined(CC_ENV_32)
const unsigned CC_MAX_NUMBER_OF_POINTS_PER_CLOUD =  128000000;
#else //CC_ENV_64 (but maybe CC_ENV_128 one day ;)
const unsigned CC_MAX_NUMBER_OF_POINTS_PER_CLOUD = 2000000000; //we must keep it below MAX_INT to avoid probable issues ;)
#endif

So the limit is 2 billion per cloud, not 4 billion (2^31 rather than 2^32)

But you can have as many clouds that size as your system can handle so if you need to look at a 14 billion point cloud, you would need to break it into 7 2 billion point clouds, currently only certain file formats support this chunking action (BIN (Since this is cloud compares format this implies opening a file in the 32 bit application which was saved from a 64 bit version), ASCII, PV, and LAS (through the PDAL plugin)).

CPU - I personally wouldn't recommend an 18 core Intel system unless you really really want Intel for some reason Intel-Core-i9-9980XE-vs-AMD-Ryzen-9-3950X. Just make sure you don't go after a threadripper looking to utilize all cores, the 2 numa nodes from threadripper will currently not both function in the vast majority of CloudCompare. So the Ryzen 9 3950X with 16 core 32 threads is likely the highest performance chip for CloudCompare. Only the Threadripper 3980X and 3990X will use more threads at once then the i9-9980XE but they would be wasting half of the available cores.
If I was building a computer for myself to day it would be hard to justify anything more expensive than the 3950X.

RAM, more Ram better right? for the most part yes, as long as you have enough to not page to and from long term storage you should be good.

As far as GPU's go, remember that CloudCompare is currently only using OpenGl version 2.1 so no hardware advantage from extra features on the card, meaning a GTX 1080 TI would likely in CC outperform an RTX 2070 or possibly even an RTX 2080. The system just won't take advantage of any tensor or ray tracing functions available.

[CloudyCloud]

Do you have a support for RTX cards in your roadmap? Would that bring any significant speed ups?

[WargodHernandez]

Nothing planned, not sure what we could utilize raytracing on GPU for to be honest. The only current raytracing used at all is in the plugin which allows you to calculate ilumanace based one Rays being cast from a hemisphere outside of the cloud to see which points are hit and which are blocked by other points. But it's a one and done calculation, not real time.

Plus openGL doesn't support raytracing so we would have to move on from openGL to vulkan before it would even be remotely approachable