Hardware Recommendations

Any advice on building a PC whose primary use would be processing things like photogrammetry?

I’m just looking for a discussion and general advice on this topic rather than exact specifics based on a fixed budget. I had a look through the forum and didn’t find anything. There’s so many different combinations of CPUs (Xeon/i5/i7) and GPUs that sometimes it’s hard to assess what is the most important and best bang for your buck aspect of each component is.

The requirements to start with: OS and hardware requirements

depends on the size you want to make.

cpu, is by far the slowest part. normally take about 80% of processing time for me.

your 980ti is a good option. unless your doing more than a few thousand photos at a time. i wouldn’t bother upgrading that. even then its last thing to upgrade.

ram, get as much as you can. depends on size of models of course. but more ram, means less likely to run into issues with bugs software. this may change with updates though.

temp drive, large ssd. it seems speed isn’t really needed so much. but low response times help a lot. and hard’s drives a terrible for it. i haven’t seen anything more 70mb/s (on an ssd raid that can do over 2gb/s) from rc’s temp drive usage. it seems mostly cpu limited.

page file. if your doing big scenes and you don’t have enough ram. then having this on ssd raid’s help. I regularly see up to around 400gb page file being used. and I have seen it up to 1.4tb, i get speeds ranging between about 200mb/s to 1400mb/s on 2 x 950 nvm (this raid can get up to 3gb/s). so running faster storage for pagefile seems to be worth it.

my system is a 6900k @ 4ghz, 128gb ram, 2 x 1080, 2 x 950, 4 x sandisk 480gb.

typical projects i run on this take about a week… when they work.

I feel lack of ram is probably biggest issue. though cpu takes 80% of the time.

be careful though i believe cpu is limited to 32-cores per license, so that cuts out most of the new xeons. which is a bit of a pity since its by far that slowest part of the process.

Thank you very much Chris, that is a very informative post. Exactly what I was looking for.

Feel we need a bit more discussion on this. I’m working on a new build at present but as OP said there is so much choice it’s hard to know where to start.

? single intel core i7 processor vs dual xeon cpu? - number of cores vs pure processor speed?
? nvidia GPU’s - GTX 1080 single vs SLI vs GTX980 single/SLI etc etc etc…???
? SSDs, sata / nvme ??
? cooling- air cooling vs water cooling?

maybe we need some specific bench-marking that suits RC. I know that other sites have similar benching for other photogrammetry applications but RC seems a bit different in it’s algorithms.

well i guess it depends on size of model :slight_smile:

if you want to process lots of photos… gpu helps.

if you want to create a model of a large area, you need lots of cpu. though this would really be better done in network.

then getting RC to use everything it has is another thing.

I’ve found with 2 x 1080 on the job with a little over 6000 photos, at the start would bounce around up 99%. then from half way through it wouldn’t get over 30% for each gpu for the rest. that took about 1.5 days to create the depth maps.

so I would probably go one fast gpu first, and less but more powerful gpu’s. and remember you can only use up to 3 gpu’s per license.

cpu for me takes a lot longer than gpu parts. i’m running 8/16 cores at 4ghz. and these 6000 photos will probably take another 5 or 6 days to run using just cpu.

so keeping in mind the 32 core limit. the best xeon you could go for without wasting money would be a 2667 v4, which is basically the same as i have in the i7 but at 3.2

otherwise i guess you could go a single xeon with more cores, or disable hyper threading.

but really for the time it takes, it would really be better just to have 20 or more pc’s running. but that’s not really feasible at least for me due to license costs.

for me though the biggest thing is max out the ram. 128gb i’m finding is nowhere near enough. i’ve seen pagefile usage over 1tb, but i think a lot of my stuff would fit in 512gb.

saying this, i may have found a solution to my constant crashes due to lack of ram. but I’ll have to wait a while longer to find out.

for storage, i have 2 raid array’s, one is a 2 x 950 nvm. the other is 4 x sandisk.

the 950’s are faster. but they are a bit more all over the place in speeds. you need the setup right, which includes formatting them with less than their full amount. and then don’t fill them up. i just use these for pagefile, thats it. (i also had 1 working at 1x at first and that dropped the speed a lot)

the 4 x sandisk, i think are a better deal. they just work. and i think space and low response times are more important than top speed. though i would be interested in seeing how intel 750 go, i have a feeling they may be better than the 950’s.

but as i said in earlier post rc dosen’t take full advantage of the ssd’s on the temp drive, so even a large single ssd would be ok.

pagefile on the other hand i’ve seen use up to 1500mb/s which is about 50% of what the array could do.

coming from doing a lot of 3d work (mostly still, but some animations too). I find doing photogrammetry a lot more processor intensive and a lot longer wait times to what I’m used to. (and i have some really long render times). as well as having only 1 pc instead of being able use our small render farm.

so i would go in order (for large scenes with current rc version and licensing, ram usage could change to need a lot less, and network processing may change the setup drastically)

Ram
Cpu
GPU
SSD

in a future rc version that included network processing for reconstruction phase i could see having a pc with a few gpu’s to make the depth maps and the rest being fast i7’s would be a better setup. but i also think RC ram usage probably also needs to get under control a bit for this to be practical.

Chris,

This is great information. Thanks for taking the time to post.

I’m persisting in my attempts to get RC to work for me as I reckon it will give me better results and much , much faster than the other big-name apps for large datasets (just need to get it to actually work, without crashing every time I forget to cross my fingers/toes etc!!)

I am currently looking at a more capable build- either:

  • intel i7 6900 (8cores), 128Mb ram, 500Gb nvme ssd (+/- sata ssd) + 4Gb HDD, GTX 1080 GPU
    or
  • dual xeon e5-2670 (16 cores) more RAM … say 512Gb depending on $$, 500Gb SSD + 4Gb HDD, GTX 1080 GPU
    (similar to here https://youtu.be/ak8GsI0bbjA

With either of these builds I get the impression you think it would be more beneficial to spend an extra $1000 ish on more RAM than adding in a second 1080??

Maybe I’ll see if I can do both builds- get the i7 one via work and gradually build me a xeon monster at home as funds allow- that would probably be a good comparison for RC setup- only issue I see is with licensing at present you have to wait min 7 days to transfer from one machine to another to test… unless I go and buy another licence I guess. Hesitant to do this at present as I have not had a successful output from RC in quite a while.

i’m looking for a new 512GB SSD,
can anyone help me to find the best one?

Hi john9871j

Samsung PRO SSD versions, not EVO…

i found this article on google
http://www.deskdecode.com/top-best-480gb-500gb-512gb-solid-state-drive-monthly-updated/
and they suggest me to buy samsung 850 evo.
how about that?

Hi john9871j
EVO vs PRO version is not about speed but durability, EVOs are rated around 100 TBW ( 100 terabytes written ) vs 200-400 TBW for PRO versions. There are even better drivers regarding durability and they can withstand approximately 1500 - 2000 TBW.

If you were an “ordinary” customer with “few” writes, I would say EVOs are OK, but if you work as a professional in the field and do a lot of data processing, then the PRO is a better alternative ( the difference in price is not so big )