Re:New Optimizations Improve Deep Learning Frameworks for CPUs
by Eridrus on Saturday October 14, 2017 @09:32PM
(#15475372)
This seems like a good study in how to present misleading benchmark numbers.
They have a 72x improvement, great! So, first start with a really crappy baseline not tuned for your high end 68 core machine - which runs slower than the 22 core machine so that you have a tonne of space to improve.
Then pump up the batch size to 2048 images per batch because it's an easy know to twiddle that has nothing to do with the implementation speed and you don't really care about the actual model accuracy and you just want to prove how fast you can make this. (I'm ignoring the recent work on making large batch training work, but the larger point is that batch size is a model hyperparameter that influences more than just speed)
And then ignore the fact that it's performance per $ that people care about, not pure speed.
Re:New Optimizations Improve Deep Learning Frameworks for CPUs
by naturalgradient on Sunday October 15, 2017 @04:40AM
(#15476404)
What a bad, outdated article. I don't understand how articles from this website keep appearing on the first page here, they are usually badly researched puff pieces.
Aside from the nonsensical performance numbers:
Theano, Neon, Torch, really?
No mention of PyTorch, Theano actually being discontinued..
Re:New Optimizations Improve Deep Learning Frameworks for CPUs
by antorsae on Monday October 16, 2017 @05:41AM
(#15481982)
And the least expensive Intel Xeon SP Platinum you can buy is ~$3000. No wonder why there's not a head-to-head benchmark on a cost basis, e.g. epochs/min per $.
They have a 72x improvement, great! So, first start with a really crappy baseline not tuned for your high end 68 core machine - which runs slower than the 22 core machine so that you have a tonne of space to improve.
Then pump up the batch size to 2048 images per batch because it's an easy know to twiddle that has nothing to do with the implementation speed and you don't really care about the actual model accuracy and you just want to prove how fast you can make this. (I'm ignoring the recent work on making large batch training work, but the larger point is that batch size is a model hyperparameter that influences more than just speed)
And then ignore the fact that it's performance per $ that people care about, not pure speed.
Aside from the nonsensical performance numbers:
Theano, Neon, Torch, really?
No mention of PyTorch, Theano actually being discontinued..