Maximum computer for time reduction.

PIX4Dmapper PIX4Dmapper Questions/Troubleshooting
1

Hello

We work in a field environment collecting aerial images and it is apparent that our current computer setup is inadequate to process (render) the 3d mosaic image in pix4d in a suitable time (typically 2+ hrs)…we need to reduce this time as much as possible. Time reduction trumps cost of a new computer system by a very significant amount.

Initially the first thing that comes to mind is a dual cpu system for a total of 16 cores. Im not sure if this is a good idea yet but thats my initial thought.

What are the software limitations…would the software be able to take advantage of a dual cpu system?

Do you have an example of a system that could render 200 photos in say 30 minutes?

here is a quick example of something similar to what I was thinking about

http://www.newegg.com/Product/Product.aspx?Item=9SIA68F21F3010&cm_re=Xeon_processor_e5--9SIA68F21F3010--Product

What are your recommendations for a pix4d workstation that can provide the ultimate pix4d rendering experience (fastest time)

Thanks

-Nick

2

Hello Nick,

This question is very interesting and many users asks us about hardware recommendations.

The most important is to have a balanced computer so there is no bottle neck.

The 4 key components from the computer are:

CPU
RAM
Hard Disk
GPU

Different steps uses the resources in different way, but at the same time, within a step, in some moments some small instructions needs to be completed, which cannot be paralleled and therefore it will not use the full resources. For more information about Hardware components usage when processing with Pix4Dmapper: https://support.pix4d.com/hc/en-us/articles/203405619

At the same time, the software is optimized for middle and good computers, the one you suggest is a super computer, the speed will be better, but for example, double speed in a computer (even in a balanced computer, and comparing a medium computer and a good computer) maybe does not mean double speed in processing, so between a super computer and a super high computer, the difference will be very low.

The biggest difference may come when processing very large datasets, for example datasets with 10.000 images or more, then a medium computer may fail, a good computer will works but it may over pass its resources so it may run will slow, and then a very super computer will be optimal.

For a project like 5000 images, a low computer may fail, a medium will works but it is over passing its resources so it will goes very slow, a good computer will be optimal and a super computer will not makes a difference or a very little one.

Since the speed is related to amount of images, resolution, number of features found,etc, to really compare it would be needed to process the same project in another computer and see the processing times difference.
For more information about the processing speed: https://support.pix4d.com/hc/en-us/articles/204191535

As minimum and recommended hardware we have:
https://support.pix4d.com/hc/en-us/articles/entries/26799527

As a good computer, we suggest in general the computers we are using ourselves, which are tested and are working very well:
i7-3930k (6/12 cores/threads), 32-64 RAM, 256GB SSD, GTX780 ti-4GB

As super computer we suggest:
CPU: Intel Xeon E5-2697 v3 @ 2.60GHz
RAM: DDR4-2400 256 GB.
SSD: 1024 GB.
GPU: Geforce TITAN BLACK or tesla, or two titan black in parallel.

For more information about Hardware and Software Configuration: https://support.pix4d.com/hc/en-us/articles/202559159
For more information about the use of the GPU: https://support.pix4d.com/hc/en-us/articles/203405619

For example, when an user really wants to increase the speed of the project, and they have very good quality images and the overlap is very good, it is possible to work with the processing options to reduce the processing time:
https://support.pix4d.com/hc/en-us/articles/20256041

In this case where the dataset is very good, reducing the processing options will not have impact or very little in the results, the processing options are optimized for a normal user, which does not use to have the optimal dataset.

At the same time it is possible to divide datasets following:
https://support.pix4d.com/hc/en-us/articles/202558579
or
https://support.pix4d.com/hc/en-us/articles/202557559#label10

For more information please check:
https://support.pix4d.com/hc/en-us/articles/202559519
https://support.pix4d.com/hc/en-us/articles/202559159

Regards,

1 Like
3

Thank you for the very helpful answer.

If I built a system to the specs you provided AND with a single titan z with 5760 cuda cores ( titan x black is discontinued i think)
what would be the approx render time on medium sized project with 200 to 500 images

do you have a motherboard recommendation that would support the xeon processor and the 256gb of ram. A board that your company has found to be very stable and very reliable (long life span)??

Thanks again

-Nick

4

Hello Nick,

The speed is very relative and depends on many factors, not only the amount of images and their resolution (as example, a good overlap produce a faster calibration, and it increase the speed since part of the calibration cannot be paralleled):

Depending the setting options, quality of dataset, terrain, etc, a 500 images project can take from 1 to 10 hours:
https://support.pix4d.com/hc/en-us/articles/204191535

Using the default processing options and a good quality dataset with 14 MP images, it can take 4-7 hours (aprox)

A Xeon with 256 ram and Geoforce Titan, and SSD, will give more speed than an CPU: Intel i7 5960K, 64 GB, Geoforce GTX 970 and SS.

Since 200-500 images is not a very large project, as larger the project as more difference in the speed.

And the tendency in the future is that the difference will be higher between a good machine a normal.

Regarding the mother board, and components, we have here some computers with:
Board: ASUS Z9PE-D8 WS (Dual CPU) (LGA 2011, Intel C602, E-ATX, XL-ATX)
CPU: Intel Xeon E5-2650v2 (LGA 2011, 2.60GHz)
GPU: ASUS GTX Titan Black (GK110, 6Go)
RAM: DDR4-2400 256 GB.
SSD: 1024 GB Crutial Mx550

Regards,

5

Very useful thread here.

Assuming that I have a good dataset with good overlap and about 2000 images. How long will it take to procude Point Cloud Densification for a workstation at the following configuration:

Motherboard - Gigabyte P85
CPU - Intel Core I7-4790 (3.6GHz)
RAM - DDR 3 8Gb Patriot x 4 = 32 Gb
GPU - Gigabyte GeForce GTX 760
SSD 1 - Patriot 120Gb
SSD 2 - Patriot 240Gb

Thanks.

6

Hello Vasil,

Considering that the geolocation is good, it may take around 20-30 hours. This value is relative since the computer may be doing some internal operations when processing that slow down the processing, or other factors involved.

When processing a very large dataset, is when a very good computer makes more difference.

Regards,

7

My workstation is dedicated for pix4d.
I have no geo location, but I will add GCPs at a later stage. Will it speed up the processing if I add GCPs at the very beginning?

8

Hello Vusal,

The GCPs do not speed up the process, they help to adjust the model better and help to place the model into the real word with more accuracy.

Regards,

9

Dear Support, I am also looking to upgrade a few components to speed up the processing as well.

 

My current specs are,

CPU: Intel i5 4460 3.25Ghz

GPU: Ge-force GTX970

MB: MSI B85-G43 Gaming

RAM:  32GB  4 x 8GB DDR3 1600

HDD x 4: 1x 2TB,  2x 1TB, 1x 500GB (Not using RAID)

OS: Win 7 Pro 64bit SP1

The two main ones I will change is to add an SDD and add a better CPU.   I have already ordered a SanDisk Extreme Pro 480GB SDD but now I cant decided which would be the better option for the CPU choices to match my current GPU/RAM and MB configuration.

The two I am looking at with a budget of between $1000 to $1500 are theIntel Core i7-5960X Processor Extreme Edition 20M 3.5G LGA2011  or theIntel Xeon Processor E7-4809 v3 (20M Cache, 2.00 GHz) 

Both have a 20MB cache and the main differences is the Xenon is 2.00Ghz vs the i7s 3.5Ghz. So I would assume the i7 would be the better choice and also considering its $200 less than the Xenon?

I cant see any advantage that this model Xenon has over the i7 to warrant paying the extra $200. I have no issue to pay more and up to $1500 if there would be some advantages. Can you please advise thoughts and recommendations on this?

Thanks in advance , Regards John

 

10

Dear John,

Unfortunately, we cannot provide an informed decision about this particular hardware: there are too many factors and possibilities and we cannot pinpoint one deciding factor. They seem quite similar and we do not expect that there would be a great difference between the two. Hence price might be a deciding factor for you.

Regards,

11

When I did a lot of video work, hard drive throughput was critical. It seems to me that with Pix4D hard drive throughput may be less of a limiting factor than CPU speed and RAM. If I’m accessing pictures off a server (which is in the same building, FWIW), am I going to see appreciably slower processing than if I’m accessing pictures directly off my computer’s disk? Thanks in advance for any feedback!

-Desmond

12

Hello!

I´m planning to upgrade my PC to use pix4d.

I would like to know if there is a difference in performance between using ddr3 or ddr4 ram?

if there is, how much is it?

13

I think this link is one of the best write ups between them.

14

Hello,

I need to process 4000 to 5000 images, Phantom 4 file 12 MP camera. Please guide us how much time it will take to process and what computer/work station specification we require. All 3 steps should be completed in almost 20 to 30 hours. We need to do this urgently.

Thank you

 

15

You need to advise a lot more about what you are doing and what you want to achieve. Is it land or crop survey or DSM, or 3D terrain for quarry or mining surveys or 3D building and structure model and surveys? Also what resolution will you want to output?. At the max resolution with that many photos I suspect will take 3-6 days at least for a 3D point cloud and mesh even on a top spec machine. At a very ultra low resolution it might tank 5-12hrs at my rough guess. Best if you can post some more details here.

16

3000 images , city reconstruction model. Resolution is medium, standard as per your 3d model template, also all 3 steps .

17

 What is the area coverage? At this size you might be best to split the project into 4 to 6 segments. You will likely need a very high end spec system to do this anywhere near your target time frame and I even doubt this will be possible with just one system.

I have done medium processing with 400 photos and it took 24-30hrs on that alone.  The CPU is critical to be as high end spec as you can afford as this does most of the work. At higher resolution its tank nearly 3 days.

This following is my specs and I suggest you also look atthe wiki for more info on the details which systems are the best for this volume of work.

CPU: Intel(R) Core™ i7-5960X CPU @ 3.00GHz
RAM: 32GB DDR4
GPU: NVIDIA GeForce GTX 970 (Driver: 10.18.13.6839), RDPDD Chained DD (Driver: unknown), RDP Encoder Mirror.

SDD Hard drive

 

Also check this page and especially the very good machine specs which I guess is a going to be a minimum what you will need to crunch that volume of data.

A very good machine:

  • CPU: Intel i7 5960X (8 cores, 16 threads).
  • RAM: DDR4-2400 64 GB.
  • SSD: Crucial MX100 512 GB.
  • GPU: GTX 980, TITAN BLACK.
18

Hello.

I am building a machine specifically for running Pix4D and have a question about the benefit of the number of GPU CUDA cores:

Is there any particular advantage between the following nVidia graphics cards, considering the difference in cost (these seem to be current replacements for those referenced elsewhere in the support pages):

GTX1050, 4 GB, 768 CUDA cores, $215

GTX1060, 6 GB,1280 CUDA cores, $380

GTX1070, 8 GB, 1920 CUDA cores, $630

Other computer specs are: Intel i7 8-core 6900K 3.2 GHz CPU with liquid cooling, 32 GB RAM, 256 GB SSD, 2x 3TB HDDs (RAID) for data storage.

The CPU is quite expensive but important, so I am trying to economize a bit by not spending more than necessary on a GPU that may not have a corresponding benefit.

My typical projects involve processing about 1000 to 2000 geo-referenced images captured with an eBee and Sony WX camera.

Thank you for the advice.

19

 FWIW, I got a new Microsoft SurfaceBook its only dual core i7 2.6 GHz, 8GB ram - I set about processing a 1000 pic model (12 MP pics) - and was disappointed to find it took 24 hours to complete.  I had to redo the model with GCP’s, this time I used a quadcore i7 (4790 @ 3.6 GHZ) desktop with 16 GB ram - so far its taken 30 hours, and is still labouring through ‘generate dsm’, with 4 steps to go after that.  Not sure if this says the surfacebook rocks, or if there is some other explanation.  I should also add, while the surfacebook has a discrete GPU, i had forgot to enable it, so that result was with intel HD500 graphics. The desktop has a Nvdia Quadro K620.

20

What resolution settings did you set it at? I am surprised you got such a good result with the SurfaceBook to process 1000x12mp photos even if it was set to lower resolution.  A 4th generation quad cor  i7 processor is probably a whole lot slower than a 6th generation i7 that is in your SB. Keep in mind a big chunk of the Pix4D processing work is mostly done by the CPU and some shared with the GPU and not so much by the ram.

My PC has 32gbDDR4 Ram and an i7-5960x CPU  which all up is 16 core and even using this , procession full resolution shoots of 300x5mp photos can take up to 10-20hrs with all 16 cores running full on at 98%