search

OpenDroneMap / ODM

A command line toolkit to generate maps, point clouds, 3D models and DEMs from drone, balloon or kite images. 📷
https://opendronemap.org
GNU Affero General Public License v3.0
4.92k stars 1.11k forks source link

Recommend Hardware? #333

Closed Bavar2142 closed 8 years ago

Bavar2142 commented 8 years ago

Hello all, apologies if this is a daft question and feel free to move to appropriate area if this isnt the place. Does anyone know if the ghz of the cpu or the number of cores is more important for ODM? We're looking at building a processing rig on this end and are thinking of 64 gig ram and a samsung 850 pro ssd.

alexhagiopol commented 8 years ago

Hi @Bavar2142 I've been working extensively with ODM and other 3D reconstruction libraries. Here's a list of parts in my development rig. Let me know if you have questions or if you're thinking of building the same one. The only changes I recommend are (1) maybe getting a cheaper GPU and (2) getting a single set of 64GB of RAM as opposed to two sets of 32GB (e.g. https://amzn.com/B00PLERJ7M).

http://pcpartpicker.com/list/sYcPsJ

Good luck!

Alex

dakotabenjamin commented 8 years ago

I don't know if I can recommend GHz over cores, but on my work computer (not the server we do production stuff on) I have an Intel Xeon CPU E5-1650 (3.5Gz) and 32 GB memory. We just installed an Nvidia 1080 (stock) but right now ODM does not support OpenCL or any GPU processing.

alexhagiopol commented 8 years ago

Just realized I didn't answer the question. I'd say GHz and cores are about equally important. That's actually why I went the i7 route and overclocked to 4.3GHz on 6 cores. Highly recommend the 850 Pro SSDs as well.

GeospatialDaryl commented 8 years ago

I'm a Samsung SSD fanboy as well - but for almost all applications, the EVOs are fine. The Pros are significantly more expensive. Unless you are doing server database transactions, it's unlikely you'd run it hard enough to see the long-term performance difference between the Pro and EVO. One technology I'd look for in a motherboard is support for NVMe boots - both in the PCIe form factor and M.2 (NVMe, not SATA) on the motherboard. Wicked fast.

While intel continues to dominate for instructions per clock, there is a substantial value (still, especially for home lab crunch boxes) in the aging AMD FX line. If you have the budget, those Xeons are sure nice - but take a close look at the FX 8320e (8 cores for ~$130). Mine overclocked to 4GHz without any fiddling - on a Corsair Hyper 212 EVO. I'm looking to get it up to 4.3 or so when I have an afternoon for that. For my applications, more cores is better than more IPC. And the MBs and CPUs are cheap these days. If you have a few months, I'd wait for AMD Zen.

Just my 2 cents-

Daryl

Bavar2142 commented 8 years ago

Hello all, thanks for all the replies. In terms of the cpu I have a 6600K on my home rig that ive gotten up to .43 ghz. Would that approach be worth a look? Hows the 1080 treating you? @dakotabenjamin

In terms of your build @alexhagiopol at the moment the changed id make if I was building today would be H110 or 115 for AIO cooler, a 700 series gpu (just to save money with the processing being cpu dependant) WHat made you recommend a single stick of ram?

Bavar2142 commented 8 years ago

In terms of time table itll be a while as we are waiting on hiring a new it person so zen may be an option.

Bavar2142 commented 8 years ago

@dakotabenjamin Do you know if there are any plans to share some of the processing load with the GPU eventually?

Fi156 commented 8 years ago

Pull request will be very welcome, but as far as I know, nothing is planned yet...

Bavar2142 commented 8 years ago

@Fi156 If I had the programming skill id love to do that. Sadly im just getting back into it after a long absence.

alexhagiopol commented 8 years ago

@Bavar2142 If you look at the amazon link I posted for the RAM, you'll see it's 8 sticks of 8GB each. It's important to get them all in the same set because they are tested and validated together at the factory.

I don't tend to recommend the all-in-one liquid coolers because they are louder and have more points of failure than a good air cooler.

EDIT: I was actually thinking of making another dev rig, so I created this cheaper and more powerful part list based on what I learned from building the first one: http://pcpartpicker.com/list/zXyXPs

dakotabenjamin commented 8 years ago

GPU processing is certainly in the future but right now we don't have any timeline. When we build and integrate semi-global matching that would be a prime candidate for GPU processing

Bavar2142 commented 8 years ago

Thnaks for the heads up @dakotabenjamin Thanks for the explanation @alexhagiopol I had originally thought you meant a single stick of ram. Ill paste a link to the parts list that we're looking at once I find it.

edgarriba commented 8 years ago

some months ago I started a GPU based approach for features detection and matching in OpenSfM. Unfortunately, my schedule has changed and I cannot spend time on this right now. Probably yes in the future. However, if someone wants to follow with that PR will be great!

https://github.com/mapillary/OpenSfM/pull/49