Oh and I love the post about the Ram. Ryzen uses infinity fabric technology to connect the two dies to the socket and each other. That technology is tied to the speed of your Ram. If you use slower Ram with Ryzen, your results fall off a cliff (10-30% in other multicore benchmarks).
Sounds a bit like apples-to-oranges, especially when fundamental architectures are mentioned. Thatās where I get lost given stats will change every 6 months or so with new hardwareās and softwareās.
Finding optimized hardwareās, softwareās and work-flows for high fidelity photogrammetric measurements is more than a full time job for me, that aside trying to produce outputs that someone may want to pay for on-top of it all!
I commend you-all for the patience, time and financial efforts put into this research.
Thanks for posting here.
@Marilli - Ā It will be nice to see some newer benchmarks fromĀ the Pix4D group.Ā
@Gary La Mantia - Im not sure if you were responding to my post of Ryzen vs. Intel benchmarks but I agree it is sort of a apples and oranges comparison. Every person is probably rendering different settings and looking for different outputs. For me and my operation this was the best way to compare the two different machines. I also did some testing on a MacPro with BootCamp that I use and I am going to test the Mac version of Pix4D on my mac pro vs my intel machine. Im sure the Intel machine is faster but I amĀ not sure how much. My mac pro is a 12 Core 3.46 Xeon with 64 GB Ram and a Nvidia 1060.
@ Sterling and Everyone else - A thought I had was we could have a thread of a specific project and include the project files. Anyone could download it and return the report to the Pix4D staff or a moderator who could compile the data.
It would be nice to have a variety of types of projects so those interested in what machine to build/buy could tailor it to their specific use.
The reason I say this is because from my findings in my own use some projects would render better with one machine part of the year then better with another later due to the image count and GPU use.
Ā
Just some ideas and I would be happy to test a project for someone if they are really looking into building a machine like mine.
FYI my current computer cost me about $3000 to build. If any one would like the part list I can link to it.
Ā
Brian
This was suggested, bench mark data sets!
Pix4d has a good diversity of data but absolutely no precise measurements to calibrate by.Ā
I like all of this talk!
I like Pix4D too. I fear without this type of talk, workflows will migrate to other ecosystem such as Bently or ESRI, with better rounded workflows.
Pix4D outputs are very good for the dollars compared to must anything else.
Yes, Iād be more then happy to crunch a few benchmarks through my current system for comparison.
@Brian Young: I like your idea of creating a post with datasets available for everyone to download and crunch through with their hardware. This would enable to create a community benchmark and would help select processing hardware. Would you be interested in creating and handling this post? I can for sure help out as well.Ā
Anyone willing to provide datasets for that purpose?
It would be good to have datasets of different sizes (nb of images, resolution,ā¦) with a good flight plan (i.e. enough overlap,ā¦). Ideally following the recommendations from that article: https://support.pix4d.com/hc/en-us/articles/202557459
Having different use cases would be great too, e.g. construction, surveying, Agā¦
In addition to processing time for the different steps and the hardware, interesting information to provide for such a benchmark would beĀ the version of the software. As we constantly work on improving the algorithms the performance can change from one version to the next. Comparing the processing using the same processing options would be good as well, e.g. define that a certain dataset should be run with the 3D Maps template. What else would you include?
Yes a bench mark data sets would be most useful,
I am also considering buying a workstation soon and right now the only relevant information I found was mostly this thread and benchmark from comparable softwares (anandtech is using Agisoft photoscan for benchmarking : http://www.anandtech.com/bench/CPU/1610, but itās an old version of the software and from what I read Pix4D is often faster than Agisoft. So the result would probably be different).
According to those benchmark, i7 7820X is the best choice if we take into account price/speed (which is a bit surprising since I was thinking a ryzen system would be the winner in this categorieā¦)
Anybody has any experience with this one ? I was waiting for threadripper but the results are not good on this benchmark, I would like to have some better data from user using Pix4D before jumping in^^.
Iām usingĀ a Threadripper (TR) 1950x. Iāve seen processing times cut down anywhere from 40-90% from a Ryzen 7. I canāt compare to an Intel chip though since my last one was a 6700K, and I wasnāt happy with it. As said the value of a particular setup seems to depend on the types of workloads you are feeding in. I really hope Pix4D does at some point begin posting some benchmarks though I can see the difficulty they will have in standardizing the tests since the MBs themselves can become a limiting factor. For example, with the 7820x you only have 24 PCI-E lanes available to GPUs and M2 drives. If you are going to SLI cards or run multiple M2ās in Raid 0 or 10 to match your high end CPU then your MB will become the bottleneck. With TR you get 60 PCI-E lanes available to cards. Iāve noticed a large speed increase with Raid 0 M2 drives on an x399 MB and Intel doesnāt support that. Iām also going to get a second 1070 to run in SLI and Iāll still have PCI-E lanes left should I need to add additional cards or drives.
That all being said if your workload prefers high processing speeds I could see why the 7820x would win a comparison. Stock it can achieve the highest throughput pure core of any of the processors in the test. It has more thermal headroom than the 7900x and more throughput per core than the 1950x. TR is roughly equivalent to Broadwell-E chips in a core on core test so anything you do that needs maximum pure core throughput the Intel Skylake-X line will win out.
That all being said Iām very happy with TR so far. It cost roughly $500 less to build my TR/X399 system than it would to build an i9 7900x system and I do other things with my PC that favor more cores like streaming my work as Iām doing it for others to learn from.
2 Likes
@ Pierangelo
I could help with the community benchmark. I have a couple of questions.Ā
-
Where could we host the files that will be used in the bench marking? The files that I referenced in my first post above are 2-6 GB each. I would be happy to let other use them for benchmarking.
-
I only really work with the Ag side of things as of right now. Is there much difference when rendering different data sets? Mining, 3D Models, Highways,ā¦
Ā
I think that this could all be done together and would like to get a feel for who else would be interested and what areas they are currently working in.
Ā
I will try and respond a little quicker in the future. We are starting to get into the harvest season here in Iowa and I have been getting our combine and tractors service to go to the fields.
Ā
Brian.
Ā
@ Sterling Mitchell
Would you be interested in running my three projects on your machine sometime in the future. I would like to see how it does before I look into a new machine build this coming spring. It does not have to be now but maybe in the next few months.Ā
I would like toĀ find out if the pix4D company would be willing to setup some sort of hosting for the benchmarking data sets
What is your current machine specs?
Ā
Brian.
Iād be interested in doing this though again Iām a little timid that possibly my GPU and/or memory speed could be bottle necks for the program. Iād hate to post times only to see a new i9 with a Titan xP or 1080Ti slant the true message. Here are my current specs though I am looking into a second 1070 in SLi or a single Volta card once they are available. If I get a Volta card I will post before and after numbers. I will also be building a M2 Raid array at some point in the future.
AMD Ryzen Threadripper 1950x (stock clock, seems to consistently use XFR during Pix4D use which clocks to 4200 mHz)
-Slight aside but Ryzen in general seems more intelligent when it comes to utilizing extra TDP headroom from the CPU cooler, if you give Ryzen a big AIO cooler it clocks well. If itās on the stock fan, good luck seeing above floor clocks without OCāing. Iāve seen this on both the Ryzen 7 and Threadripper (Skylake i7 and Skylake-X i9 priced chips)
Gigabyte Gaming 7 X399 MB
Kingston HyperX 2666 mhz quad channel memory (4x8 GB)
Asus Strix nVidia GTX 1070 (8 GB DDR5 and stock clock)
Samsung 960 Evo 1 TB (M2 drive)
Ā
Ā
Ā
@Brian: I can provide access to a Dropbox folderĀ to store the projects. For now the system that I think will work best is to create a āFile requestā on Dropbox, where people that are interested can upload the images of their project (as a .zip file if possible). Then, I can move these files to a shared Dropbox folder, so that everyone can access and download the images. This makes sure that everyone has download rights, but not āEditā rights, so that the datasets are not deleted by mistake (or on purpose). Another advantage is that the datasets will be stored at the same place, so that they can be easily accessed. I can provide these links as soon as you have created the post. I suggest we name the post something like: āComputer hardware community benchmarkā, what do you think?Ā
As for the second question, there can be differences in processing times for different types of projects. There are many parameters that influence this. For example, the image content and processing options can make a difference in how an Ag project is processed in comparison to a mining project.
Sounds good. Ill get started with the post.
1 Like
@everyone: Brian has created the Pix4DĀ Computer hardware community benchmark:Ā https://community.pix4d.com/t/4079
Thanks Brian!
Feel free to upload datasets, process the ones that are available and share your results! Looking forward toĀ see how the different machines perform.Ā
As soon as an Agricultural Image Dataset is available Iād be happy to benchmark.
Ā
@Brian Young do you have the dataset you used to post the earlier Ryzen 7 1800x and Skylake i7 7700k benchmarks? Iād be curious to see what happens with a slower GPU but extremely multithreaded processor on the same standard. If you do let me know what settings you used as well to process the data (IE Ag Multispectral preset or custom)
1 Like
Hello,
Im using 1950x threadripper with x399 MB and cant get Pix4d to use the whole cpu. any ideas what I may have set wrong?
I did a test project and only seeing 20% on cpu usage.
Troy, please see your other post with my answerā¦nothing is wrong.