AlphaGo and custom hardware?

[xela]

At the time of the AlphaGo-Lee Sedol match, I remember hearing a lot about the algorithms, and the fact that computer go had achieved a new level by using smarter software, not by brute-forcing the problem "Deep Blue style". The deepmind web site mentioned that AlphaGo already plays at a very high level on an ordinary PC, and giving it a few extra cores is just the icing on the cake (details in their Nature paper (PDF download)). But since the source code hasn't been released, it's not possible to verify this.

Recently, I've been hearing about Google's new Tensor Processing Unit. They claim it was used for the AlphaGo match. I can't find any mention of this on the DeepMind web site itself, and I don't recall any mention of custom hardware during the match. Seems strange that the topic didn't come up at the time, or am I misremembering things?

So this leaves me wondering: how much of AlphaGo's success really is due to the new deep learning methods, and how much is from souped-up hardware? Thoughts on this?

[pookpooi]

I think we can use reflection of Zen, in computer go server. Zen-1c-2.8G (the version before applied deep learning) get 2979 elo by using one core of an i7 5960X.
Right now the latest version Zen-14.6-1c0g get 3748 elo (also use one core), and if add one GPU then it'll get 4115 elo.

[djhbrown]

Code: Select all

thought      1       0      0      0       0       0      0
dcnn+RL      0       0      0      1       1       1      1
monte        0       1      1      0       1       1      1
pus          2      24      1      ?       2       ?     2000
strength   Atlas    6d     3d     1d      8d     10d     10p+
example    Swim    Zen19   CS     idr   Leela   JueYi    $dog

idr = i dont remember its name, but there once was a 1d bot on kgs claiming to be pure dcnn (can't remember for sure, but also think Demis said in 2015 that Alphago without monte was about 1d)
pus = cpus+gpus+tpus+...
numbers may not be exact, but close enough to show the picture

PS telly says BOT is leading Formula 1 grand prix, so there's another sport down the drain

[Krama]

What is bothering me why won't they release new zen versions?
People have paid for their software and they should release updates.
This is a very bad sales model.

[djhbrown]

deleted by author

[Mike Novack]

At the time of the AlphaGo-Lee Sedol match, I remember hearing a lot about the algorithms, and the fact that computer go had achieved a new level by using smarter software............ and I don't recall any mention of custom hardware during the match. Seems strange that the topic didn't come up at the time, or am I misremembering things?

So this leaves me wondering: how much of AlphaGo's success really is due to the new deep learning methods, and how much is from souped-up hardware? Thoughts on this?

I'll try to explain.

A "neural net" is an ABSTRACT concept. You brain cells connect with each other to function as a neural net. In this case we have a computer running a program which emulates (implements) a neural net. That emulation could be done by a stepwise process, but since a neural net has a whole lot of "cells" each doing the same thing at the same time (receiving data from neighbors, sending data to neighbors) a parallel processor can speed things up.

Understand? The PROGRAM is implementing the neural net. The neural net is being trained (learning) to evaluate a function (given a state of the board, return a legal move, ideally the best move). That training is NOT changing the program, just data (cell values).

The IDEAL "special hardware" for implementing neural nets does not yet exist. It may never exist if it turns out to be more costly and slower than emulation using fast, more general purpose, parallel processors.

[johan-sweden]

It's described in this paper
"In-Datacenter Performance Analysis of a Tensor Processing Unit"
https://arxiv.org/ftp/arxiv/papers/1704/1704.04760.pdf

If I remember correctly the CNN1 mentioned in the paper is Alphago.