WebODM ideal computer hardware or preferred characteristics

WebODM
#1

Hi Everyone,

I purchased the windows WebODM. I have been performing tests for the past two weeks between Drone reploy, Pix4D, Maps made easy, and now WebODM.

I really don’t know very much about WebODM except it is open source which is something I know I like to support.

In any case, this post is to ask what are the most ideal hardware features that WebODM relies on the most, when you process projects on your own computer.

I have 2 computers, a desktop and a laptop.

Laptop is a 6 core 12 thread Intel i79750H with 16gigs of ram, 512gig see and an rtx2060

My desktop is a bit older, using a Delided with liquid metal Intel i97980xe with 18 cores 36 threads overclocked on all cores at 4.5ghz, 128 gig of ram, 6 tb total of 980pro SSD, and AMD RX6900XT liquid devil, it is liquid cooled, both cpu and GPU.

So today I processed exactly the same orthomosaic map project on both computers, 150 photos, ultra quality no photo resizing, just the best it can be basically.

The laptop took 5.5 hours and my desktop took 6.5 hours.

Unfortunately as you all can tell, it sucks big time that it seems not anyone anywhere in the world has done anything so any mapping software can support AMD GPUs, and there that’s my first and only beef with all the developers of mapping softwares.

Back to topic,

I used cpu tempcore to monitor the utilization of the cores and gpuz to monitor the utilization of the GPU.

My question to everyone is, why is it that for LONG periods of time, WebODM seems to be doing nothing with the CPU and the GPU? This is true even with the Nvidia card.

It’s just long long periods of time that the computer just idles, the CPU the GPU the disk no significant utilization, and then magically after some time all the cores and the GPU kick in and start processing for a little while, then again they stop and idle for long long time and the cycle repeats. In the window where you can see the process it does show text moving upward, but the CPU is less than 8% GPU is at 1% and the SSD is at 0% utilization. It just so bizarre.

Comparing the processing of projects with Pix4D, the second i press start on Pix4D, the computer runs 100% the entire time the project is processing.

Now that I mention it, I will do a comparison and create the same project on Pix4D and see how long it takes to complete using my laptop. Unfortunately pix4D refuses to properly work without a stupid Nvidia card.

So I’m just a little perplexed as to why WedODM behaves the way it does and what is the type of hardware resource that is relies most on, is it the highest CPU clock speed? The most ram, the fastest GPU?
And then why during processing it takes these long breaks and leaves the computer just idling.

Obviously the number of cores is not the most important otherwise my desktop would have had a faster time than my laptop with just One third of the number of cores. And it really isn’t the ram size cuz again my desktop was slower, so what exactly makes WebODM tick? What should our system be heavy on in order to speed up times and is it possible for WebODM to work the entire time and not take such long breaks during processing?

Thank you all! :laughing:

2 Likes
#2

Welcome!

To address the variable resource usage, in short, pretty normal:
https://docs.opendronemap.org/faq/#the-software-is-not-using-all-of-my-cpu-cores-memory-gpu-all-of-the-times-is-it-normal

As for processing time, the desktop CPU should be about 3x as fast in multithreaded workloads (a significant portion, though not all of what we do). Are you positive it isn’t thermal-throttling due to the overclock?
https://www.cpubenchmark.net/compare/3425vs3092/Intel-i7-9750H-vs-Intel-i9-7980XE

AMD and Intel (and VIA too, but they get a pass for being very small) never really released anything that competed with CUDA, and CUDA, unfortunately, was pretty unique and powerful at the time. Maybe we’ll see something happen in the Vulkan space now, or perhaps with Intel’s open Cyclomatic framework that dynamically recompiles CUDA to OpenCL.

1 Like
#3

I have this same behavior, my 8 core laptop smokes my 20 core (dual cpu) desktop with double the ram as the laptop. The time difference is almost double. Makes no sense to me. If I had know this I would not have purchased the desktop computer.

I’m running the native Windows Version on both of my computers. The laptop is running Windows 10 and the desktop is running Windows 11.

1 Like
#4

You can try the docker version. I run windows version the same time, most of the projects never finished.
My docker version is 10x times faster than window version. Doesn’t know why, but I just keep using docker version.

1 Like
#5

Hi,

Can you elaborate on what is the docker version?

Look at the photo below, only 741 photos, and already over 24 hours, and the cpu cores are barely being used at all… I don’t understand how this software is working, but there are some abysmal inefficiencies going on.

Anyone has suggestions?

Thanks

20230202_001959
20230202_0019591920×1440 192 KB

1 Like
#6

If there is little CPU activity for a larger dataset that normally indicates you’re out of physical memory (RAM) and are using SWAP/pagefile, which are orders of magnitude (hundreds to thousands of times) slower than working in physical memory.

#7

Hi thanks for your input.

That is not the case, the max ram that has been used is up to 48 gigs of ram, i have 64.

In any case, it will behave exactly the same whether I try it on this machine which is my desktop or my laptop which only has 16 gigs of ram. So i am rather confident it is not due to lack of ram that the program just sits there doing nothing. It has now been 34 hours thus far and finally now all 18 cores 36 threads are running at 100%. It’s just super weird and bizarre behavior from the program.

1 Like
#8

Check the disk available, sometime hard drive runs out of space.

#9

yeah, but I have 4 Terabytes of space in total, only using 1 terabyte, 3 free.

The project just finished, man, for only 741 photos for othromosaic only, not even 3D modeling, it took almost 56 hours…this speed is horrendous.

Map does look really good though, I suppose it is a lot cheaper than buying a Dronedeploy subscription, LoL

How do you export the map to get it ready for delivery to client? I don’t see an export option anywhere in the software.

Thanks!

1 Like
#10

The map is in the all.zip,or you can use the Download button below the Task to choose an individual product.

1 Like
#11

Ohhh, got it, thank you very much, I just saw the Download Assets tab, perfect, thank you once again!

1 Like
#12

Does the arthomosaic map always lose so much resolution when saved as JPG?

I downloaded the Tiff map, and that has the full resolution, but the JPG version seems to not be saved as full resolution.

Is it possible to save the JPG with full resolution in WebODM?

Thanks!

1 Like
#13

We default to 5cm/px for --orthophoto-resolution. You can set it to your survey GSD or finer, but be aware it carries a sometimes significant memory consumption cost while processing.

You can edit your Task options to change this, but it will need to be re-processed.

As for TIFF vs JPEG, JPEG has very small (in Geospatial terms) limits on pixel count and maximum dimensions, which can easily be exceeded, so yes, sometimes the JPEG must be a lower spatial resolution. PNG is a bit more forgiving in this regard, but not much, and will need a Worldfile to be loaded georeferenced.

The Cloud Optimized GeoTIFF we generate by default is the proper interchange format to use and share. The JPEG is really only suitable for preview purposes.

1 Like
#14

Understood, thank you very much for the detailed explanation! :slightly_smiling_face:

1 Like
#15

Don’t use ultra option next time.

2 Likes
#16

Hi, could you elaborate more on that?

When I’m providing the deliverables to my clients i would like to present them with the highest resolution maps and 3D models. That is why I am selecting ultra resolution.

1 Like
#17

Ask your customers what they require. You are potentially unnecessarily adding days to your compute time and exponentially increasing your hardware requirements to generate data that the customer doesn’t really need.

1 Like
#18

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.