Hardware recommendations for building a photogrammetry workstation

[brad3378]

I'm interested in building a new PC for processing models in 3DF Zephyr Pro.
I'm not a gamer and don't plan on using this PC for any CAD design. Only for processing photogrammetry models and I prefer to optimize it specifically for use with 3DF Zephyr Pro.
I'm budgeting about $3000.00 for this build.

Obviously, Solid State drives are desirable as well as high CPU clock speeds, and many gigs of fast RAM, but what about processor cores and video cards?

If I build a PC using a 16 core AMD Ryzen Threadripper processor running at 4.0 Ghz, will all of the cores be used? Would it be smarter to use a processor with fewer cores but a faster clock speed?

Hypothetically speaking, when it comes to video cards, would it be smarter to buy one really fast $1000 card, two $500 cards, or four $250 cards?

Is a Solid State Drive RAID worth considering? Or would it be smarter to just buy extra RAM?

[Andrea Alessi]

Hi Brad!

most of the core pipeline takes full advantage of multicore processors, however there are a few steps (for example, the clustering phase at the end of the SfM phase) that are by nature mono-core. Personally, i wouldn't worry too much about the CPU as long as you are buying a >= mid/high segment CPU. I'll admit that we have many other post processing tool that are not optimized for multi core, for example the frames extraction tool to name one (and one where i'm directly the culprit ^^' ) so in those specific tools a faster cpu with less cores would perform better- but that's again, tools, not the core pipeline.

Obviously, the more ram the better =) i personally am happy with 16GB but i mostly do small models, so given your budget i'll start with 32GB and consider may 64GB.

Zephyr uses graphic cards completely independently and i suggest more cheaper cards rather than one super expensive card. The MvS phase is the most intensive one, and it's completely run in parallel, so in layman terms, twice the cards will take half the time as we scale up linearly. The only thing you'd have to remember is that also video memory have a very important role, because if you are using high resolution images, there might not be enough ram on the video card so zephyr would have to roll back to run into cpu mode.

I would personally prefer two 500$ cards rather than 250$ cards, but if you are running for example A LOT of small images, then 4x250 would be better than 2x500 ( i.e. 2x gtx1060 4GB. We have a couple here and they perform very well!). Please also note that zephyr does not require all cards to be the same type, nor they have to be SLI, so another road we often see is to start with a high end card and then other are being added as time goes on, as price decrease overtime.

Our internal workstation all run on single SSD. I have to be honest we have not done any raid testing, but i don't expect it to create a significant difference. If anybody have done some benchmarking with raid SSD that would be interesting to me as well

[brad3378]

Andrea,
Thank you for the excellent response. I am going to try to educate myself on graphics cards before I make a final decision.

While researching GPUs, I discovered that Google and other companies are offering cloud computing services for very reasonable prices.
More information here: https://cloud.google.com/gpu/

Google offers cloud computing services with up to 8 GPUs - with hardware choices including the NVIDIA Tesla P100 which is currently selling for over $7,000 each!.
Even if I built an amazing $5000 workstation, it wouldn't even come close to the computing power of eight $7,000 GPUs and it would still eventually become obsolete while the cloud computing services would continually offer the latest hardware choices.

Do you think 3DFlow will ever partner with Google or any other cloud computing services? Or should I just build my own workstation?

[Andrea Alessi]

You're welcome =)

Zephyr can run into virtualized environment very well, in case that's the path you'd like to take i have never tried the google service, but we sometimes use some amazon EC2 instances when all our workstations are full

Just remember if you do so that you can not use microsoft's RDP protocol to connect remotely, as it does not support opengl, so you'll have to use other remote viewers such as splashtop, logmein, etc.

We are considering making a videotutorial for this specific installation for our 3Dflow academy series, but it's probably a little down the path, so if you have specific questions feel free to ask

[Nedo]

You're welcome =)
We are considering making a videotutorial for this specific installation for our 3Dflow academy series, but it's probably a little down the path, so if you have specific questions feel free to ask

Thank you for the Information Andrea, i'm also interested in cloud computing, never tried before tho. A video tutorial for cloud computing would be very nice!

Nedo

[Reuben]

I have just started using paperspace, can rent a high performance machine on a cloud computer with a dedicated GPU. Easy to start a new machine and cheap. I am running a project with 6000 images (260 hectares of terrain), gonna take a few days but moving faster than my work PC!