Thank you Chris, yes also using WebODM for my work. Tried the Lightning server but upload speeds in our neck of the woods are appaling. Also looking to invest into hardware as we seek to incorporate AI programs into the monitoring performed on the reserve. Anyways that is some impressive performance your computer churns out! Would you be willing to share the specs of your CPU and GPU? That would provide us great with some great guidance on what we'll need to get ourselves thank you.

Our HPC was a £2m grant win for our institute of zoology, what we have is 

• 8 x NVIDIA H100 GPUs that provide 640 GB total GPU memory
• 2 x Intel Xeon 8480C PCIe Gen5 CPUs with 56 cores each 2.0/2.9/3.8 GHz (base/all core turbo/Max turbo)
• 2 Tb RAM

I found that despite using the GPU version of odm (using the docker package) the processing was most intensive on the CPUs and regular RAM. I'm dubious about the value of investing in GPUs specifically for this task when you might be better off focussing on CPUs & RAM.

Real nice hardware! No doubt comes with real nice power bills I expect 😜

Quite rightly observed. I've seen this when using openAI whisper. Yes, it uses the GPU, but the CPU was hammered as well. I recently bought an RTX4090 (Not even in the same ballpark) but I did also equipment it with 128GB ram and 22 i9 cores (Around 32 hyperthreads) and enough SSD as well.

You do seem to have plenty of normal ram as well I see. You have lots of GPU power there, is this then using nvlink to talk to the GPUs? I've never seen H100's in real life.

Okay great thank you Chris, appreciate the guidance!

Been reading through the WebODM forum too and users there have echoed the same sentiments as yours too for the orthomosaic reconstruction. For the multispectral reconstruction users do also mention a GPU with some horsepower and VRAM (is that what you refer to as GPU memory?) is necessary for running the calculations that get processed into the final overlay, but certainly will look to invest in something with good processing power and high RAM capacity. 

Thanks for sharing this with me!

So Chris, this is the hardware you mean in your profile description with "I am a benthic ecologist using low-cost tech to explore..." ? 😜

@kimhendrikse fair point, I'm in a privileged position to have access to such a great server. However, I am using low cost tech for collection of data (e.g. examining deep-sea benthic habitats with a £3k camera rig - rather than £5m ROV) and the drone imagery mentioned above was from a 2nd hand UAV from ebay.

Hi William, 

I have never tried to build orthomosaic with UAV images because I have been working at large scale: Regional, National or Global level rather than local level.

But I have heard good things of Pix4D, PCI Geomatica (free trial version) and Orfeo Toolbox (https://www.orfeo-toolbox.org/CookBook/Applications/app_Mosaic.html) which is free and open source.

Many years ago I have been using the AutoStich and Image Composite Editor which are for camera pictures rather than UAV image

• https://mattabrown.github.io/autostitch.html

• https://www.microsoft.com/en-us/research/project/image-composite-editor/

  There were quite good softwares, but they might not been suitable for orthomosaic with big images.

If you want to use WebODM purchasing the correct computer, the owner of the software might recommend you the best. You might check their website or just send an email according to your needs.

Good luck.

Hello Paul

Thanks for your reply, always great to hear from a fellow Matie! We have explored Web based solutions (WebODM has a server-based processing node too), but network speeds in our neck of the woods are erratic at times which impairs large file uploads. We also look to invest in hardware to also run AI-based programs for monitoring and geospatial analysis. For my line of work a decently strong GPU with high VRAM and memory bandwidths is essential for illustrating all the various calculations that go into creating an NDVI of a large dataset, but that said also large RAM capacity and most importantly a capable CPU to string it all together. Which SA-based company did you source your hardware through? Otherwise thanks for sharing your advice, cheers!

Hi there Michael

Thanks a stack! Been looking for a discrete outline of appropriate hardware to accomplish my tasks. Am looking for something that can process 3-4ha at a time, but have the potential to process larger datasets, at a spacial resolution of 1.5cm. Thanks for sharing those specs, and moreover the performance such hardware is capable of.

Thank you David, a phenomenal source!