Jetson TX1 and Theano


I wanted to do some neural net acceleration with the Jetson TX 1 board. According the documentation, Theano (along with Caffe and Torch) are supported APIs.

However, I’m discovering that by default, Theano doesn’t use multiple GPU’s, but that special steps are needed to make this work:

I was wondering if anyone had any experience making this work on the Jetson TX 1 (i.e., were they able to employ all 192 GPUs on something)? Also, if anyone has experience, are other frameworks like Caffe and Torch better in their support for multiple GPU usage? Finally, because I know that neural net programming is a common application for the Jetson TX 1 board, is there some other way entirely that I should be going about this?

Thanks so much

The Jetson TX1 has 1 GPU which consists of 256 CUDA capable cores. The term Multiple GPUs tends to mean multiple graphics cards (GPUs) installed into a single PC or distributed cluster.