Building a server for Big Data Processing - Build for CPU or GPU Power?

[JonEllinger]

My team and I looking for advice from IT professionals and software engineers in the remote sensing industry on how to best build out a server to handle the processing and analytics of large data sets, both imagery and LiDAR.  We are looking to create point clouds from imagery as well as fuse LiDAR and imagery datasets together for analysis.  We hope to run the majority of our processing and analytics from the server over a fast connection within the office. 

We are looking for advice on how to instruct the firm who is building the server for us.  We have heard the CPU power is maily usefuly for the routine calculations of business software, but GPU power is better for calculations that involve trigonometry such as aerial photo triangulation and point cloud creation.  And that some software firms are developing their software suites to harness the many cores of GPUs rather than CPUs. 

Thoughts?  Suggestions?  Advice?  All are greatly appreciated.  jellinger722@gmail.com.

[Wishgranter]

Hi JonElinger
[/size]
what sort of datasets are you planing to process ? [/color][/size]how many images, resolution and type of reconstrucion subject are you planing ? 
 

read email.... [/color]

[hanparker]

HI Jon,

I'm using the Amazon EC2 - GPU g2.8xlarge https://aws.amazon.com/ec2/instance-types/ . Be interested to see other peoples opinions.

[JonEllinger]

Thanks for the replies.  We will be working with a broad range of imagery sets.  From UltraCam Eagle (260 megapixel, 1gb Tiffs) to much smaller Sony A7RII mirrorless camera raw files from drone platforms.  We are going to be creating 3D point clouds from the imagery and would like suggestions on how to build out a server for large data 3D model creation.  We plan to host the data on the server and are thinking about doing multicore processing on the server with GPUs for faster orthophoto creation. 

As I understand it for BIG single processes like running Aerial Triangulation and 3D point cloud triangulation and creation we might want to prefer using a very powerful CPU rather than a bunch of GPUs for those processes.  We hope to have several powerful workstations for these big processes but the rest of the time for QC and analysis we hope to have the data live on the server with analysts and technicians on virtual machines on the processing floor accessing the data on the server for their work.  We will have a 10gig line installed in the office so hopefully the transfer speed between the workstations and the server won't be an issue.   

Obviously, I'm not an IT expert so I'd love some general advice from people who have great setups at their jobs on how to design ours?  I would greatly appreciate it.