A choice of PC setup for best Nengo OpenCL simulation performance


#1

I’m graduate student from Taiwan. In our lab, we aim to do some research about cognition, auditory system and related neural model of human body. It’s now in an starting phase, so we need to setup a PC that is powerful to run the neural simulation as fast as possible. The goal is to able to process auditory signal in real-time.

I had read a way to speed up the simulation by utilizing openCL computation. And I have checked the related documentation, but not seen more detailed descriptions about suggested hardware requirements.

So I started searching on the internet, here are some questions I’m really confused about:

  1. Since nvidia recently got a powerful graphics card (geforce 1080 or so), AMD’s most recent product RX 480 series support more advanced OpenCL platform. What’s your choice or suggestions?
  2. Do Nengo OpenCL simulator utilize two graphics cards on PC? Or only one gpu is used at a time? (I’m wondering the performance of two AMD’s latest graphics cards installed in one system.)

#2

Since you’re doing stuff with the auditory system, have you already checked out Trevor Bekolay’s PhD thesis “Biologically inspired methods in speech recognition and synthesis: closing the loop”?

@Eric (the Nengo OpenCL dev) can confirm this, but I’m 75% sure that Nengo OpenCL dev only uses 1 GPU and that we generally go with NVidia for our own hardware, since we don’t use an advanced version of OpenCL (but I’m not sure which version we did use).


#3

Hello Seanny123,

Thanks for your advice and sharing of your experiences. I will go with the one powerful GPU approach. :slight_smile:
As for Trevor Bekolay’s PhD thesis, a really valuable work, is already on the study list of our group.

Dear @Eric, any suggestion for my build? Kindly give us your advice please.

Thanks again,

Edward Chen


#4

Nengo OCL does not support multiple GPUs yet, so the one powerful GPU approach sounds good.

We’ve typically used NVIDIA GPUs in the past, so they’re definitely more tested. However, we chose to use OpenCL instead of CUDA so that we could target AMD as well, and the first versions of Nengo OCL ran well on AMD (we haven’t tested more recently). So if you’d prefer AMD because they’re better value, I think you should be all right. And if you do run into any issues, I’d be happy to help debug them.


#5

Hi Eric,

Great, we’re going to go with NVIDIA GPU, it’s more powerful in one card.

Your advice really help me a lot, nice to meet you here!

Edward Chen


#6

Hi Eric,

Does Nengo OCL still does not support multiple GPUs?

Regards,
Saurabh