Processing time and processing time comparison Intel vs AMD CPUs

[cherrmax]

Processing a relatively large dataset (around 1000 images) a couple of week ago, I occasionally came across an interesting observation how much time difference it takes to process it on Intel and AMD CPU.
Just sharing the results here.

Just for experiment I used two different systems with only difference is CPU used and probably motherboards they were inserted.
I hire computer power on Hetzner Cloud I love their clear and ergonomic interface, German efficiency and convenient hourly rates compared to monthly charges on other platforms. So I can use a monstrous dedicated server, pre-configured for ODM, with 48 vCPU 192Gb RAM for 0.72 EUR an hour for a few hours to time efficiently digest a huge dataset, then delete the server until I need it again and not paying it’s 400 EUR monthly rate… I usually use less powerful and expensive option though…
I also find something like AWS or Digital Ocean are either too expensive or overloaded with options and features I don’t need… but that’s different topic to discuss I guess…

Anyway…
Before I used to always use Intel CPU as I thought there was not much difference but it seems it is and quite dramatic for specific tasks I guess.

So there are two systems I used to process my 1k images
Intel Xeon Gold, 16 vCPU, 64Gb RAM, 360GB SSD
AMD Milan Epyc 7003, 16 vCPU, 64Gb RAM, 360GB SSD

The processing parameters were the following
docker run -ti --rm -v /root/datasets:/datasets opendronemap/odm --project-path /datasets LNA --min-num-features 18000 --texturing-data-term area --fast-orthophoto --ignore-gsd

And bellow some print screens from the report files

for Intel Xeon Gold

for AMD Milan Epyc 7003

The time difference in processing of the same dataset looks quite dramatic to me
5 hours on AMD against 19 hours for Intel

And bellow is a side product of mapping activities just to decorate this boring post

[probably-not-porter]

I’ve had a similar experience, however, it doesn’t seem to be a question of AMD vs Intel. Our modest cluster with AMD EPYC 7352 CPUs (48 cores per node on 6 compute machines) far outpreforms any of our other machines, AMD or Intel. My current theory is that the relatively large L3 cache on the Epyc series makes image meshing more efficient.

Cool imagery!

[cherrmax]

Surely that’s not CPU itself, but something in them. Something like the larger size of L3 cache as you mentioned or something else makes the difference in specific calculation tasks. Nevertheless it makes my choice of AMD CPUs for processing maps quite obvious for now as it does save $ when I use it in a single non-cluster machine.