Re:Intel Announces Knights Mill: A Xeon Phi for Deep Learning
by jlouis on Thursday August 18, 2016 @06:31AM
(#12311460)
To elaborate: Tensorflow is clever because it takes your flow graph and cuts it into pieces which it then executes on computational units. Any computational unit will do, really as long as you have a backend for it. This choice makes adaptation to new system much easier.
I believe Caffe and Theano uses the same model, but I didn't study it. There are also some similarity in the model to what OCaml does with the incremental library, though it is not for machine learning.
Re:Intel Announces Knights Mill: A Xeon Phi for Deep Learning
by cs702 on Thursday August 18, 2016 @06:01PM
(#12316644)
"Granted, you do have to bother to make a fast x86 / AVX-512 port of your code."
That is exactly the problem: no one has bothered to make a fast x86 / AVX-512 port for any of the most popular frameworks (at the upstream level, not in some fork), and no one has an incentive to bother, other than Intel. For example, as far as I know, none of the popular frameworks take advantage of Intel's MKL out of the box.
Right now, if you want out-of-the-box high performance, Nvidia hardware is your only practical choice.
Granted, you do have to bother to make a fast x86 / AVX-512 port of your code. But because the shape of GPUs is so different than CPUs -- GPUs have a more complicated memory hierarchy for one thing -- I kind of doubt that "just run your CUDA code in the Phi" is going to work well for nontrivial examples.
(Disclaimer, I work at Google on CUDA support in clang. Which is awesome, you should try it out. :) Google "cuda clang" for instructions.)