12-10-2014 10:39 AM
Is there is any solution offered by NI, which enables to create similar Real-Time High-Performance Computing system but with the use of workstation not server equipment ? So I mean to use instead of server GPU Tesla S1070 (like it is in the example: "http://www.ni.com/white-paper/11972/en/") for example Tesla K20 which is a GPU card for workstation solutions. My question is actually, if there is hardware possibility to connect NI PXI to standard PC with GPU card by PCI Express interface and to use this card to support Real Time Computing.
Thank you in advance
12-10-2014 12:00 PM
Yes. There are multiple options that have additional configurations to the one referenced here. Here's a presentation I gave at NIWeek 2012 on this topic: https://decibel.Using LabVIEW and GPUs in Real-Time High-Performance Computing Systems. It does address your hardware question but its content is more current.
New MXI interfaces (e.g. MXI 8384) allow external enclosures containing GPUs to be controlled from a PXI controller directly. Prior to this, a workstation was needed because the MXI interface (8388/8389 and 8381) had to be installed in the controller slot of the chassis.
These configurations can be mixed and matched based on your needs. I have tested One Stop System's CUBE2 and CUBE3 enclosures for GPUs and successfully connected them to PXI systems with excellent results.
12-12-2014 08:12 AM
Thanks a lot for your answer . So technically, If I understood correctly, I should be able to control from my PXI system, using a PXIe-8384, the GPU on the workstation with the PCIe-8381 ?
Also, do you know if the link is transparent? I mean - would it be detected as a local resource on my PXI system?
12-12-2014 10:04 AM
I've never tried that configuration. If you are connecting to a GPU in a PC, then you would normally connect to an MXI 8381 in Slot 1 versus an MXI 8384 in another slot. This *may* work if no PXI controller is in the chassis. If not, the 8381 in slot 1 would definitely work.
The key advantage of the MXI 8384 is that you can use a GPU in an enclosure (not a workstation) and have it controlled by a PXI controller in the chassis. Also, a chassis may have multiple 8384s connecting to multiple PCIe enclosures - all of which contain GPUs accessible to the PXI controller for co-processing.
This is a different configuraiton option than has been historically available using the MXI technology so it can take some time to digest the options. Unfortunately, I don't have a document which a specific architecture that explores this.
However, I did give a talk at NVIDIA's GTC 2014 - Developing a System For Real-Time Numerical Simulation During Physical Experiments in a Wave Propaga... - which presents an application where this type of platform could be used. See slide 56.