Intell 48-core CPU scales to 1k cores

[jshriver]

Thought this was interesting and wanted to post here.

http://www.goodgearguide.com.au/article ... _possible/

-Josh

[Milos]

Thought this was interesting and wanted to post here.

http://www.goodgearguide.com.au/article ... _possible/

The whole experiment is a cheat. It's not sustainable from the point of heat dissipation of the package.
They've built 48 cores and said, oh we used old '94 Pentium because performance is not important.
The truth is, they would never achieve 25-125W dissipation if they used today's i7 cores with 45nm technology. The total dissipation would easily go over 1kW and that's hardly coolable even with liquid nitrogen.
And even if they managed to make something like that working its life span would be ridiculously short.

[Vinvin]

A dream for go engines or other Monte Carlo based programs ...

Thought this was interesting and wanted to post here.

http://www.goodgearguide.com.au/article ... _possible/

-Josh

[Gian-Carlo Pascutto]

A dream for go engines or other Monte Carlo based programs ...

Why the illusion this is any better fit for go than for chess?

[Gian-Carlo Pascutto]

The whole experiment is a cheat. It's not sustainable from the point of heat dissipation of the package.
They've built 48 cores and said, oh we used old '94 Pentium because performance is not important.

The experiment is a cheat? So, the 48 core chip doesn't really exist? Or what are you claiming?

The truth is, they would never achieve 25-125W dissipation if they used today's i7 cores with 45nm technology.

Of course not. Those cores are designed to achieve outstanding single-thread performance, rather than the most optimal size/dissipation/nr cores/core performance product.

[Milos]

The experiment is a cheat? So, the 48 core chip doesn't really exist? Or what are you claiming?

Are you unable to understand what's written or just playing stupid?
The cheat is to use '94 Pentium as a core and claim 1000 cores scaling without problem. Dissipation is a problem without solution. Ignoring it just doesn't make it go away. If you don't understand that, maybe you should go back to basics.

Of course not. Those cores are designed to achieve outstanding single-thread performance, rather than the most optimal size/dissipation/nr cores/core performance product.

This is just BS. What kind of outstanding performance on the level of '94 Pentium are you talking about???
The cores are not designed for anything specific. They just took the most advanced core that would still allow them to pack 48 cores without burning the package.

[Vinvin]

A dream for go engines or other Monte Carlo based programs ...

Why the illusion this is any better fit for go than for chess?

Because Monte Carlo is more efficient with a lot of (>50) cores.

[Gian-Carlo Pascutto]

A dream for go engines or other Monte Carlo based programs ...

Why the illusion this is any better fit for go than for chess?

Because Monte Carlo is more efficient with a lot of (>50) cores.

That is a bit like saying minimax is more efficient with 50 cores than alpha-beta.

Good programs use a tree search on top of Monte Carlo, and scaling that search is as hard for Go as it is for chess.

[rbarreira]

Good programs use a tree search on top of Monte Carlo, and scaling that search is as hard for Go as it is for chess.

But the average branching factor is higher in Go, correct? Around 200 instead of 35 or so I've heard. So it should be easier to parallelize a Go tree.

[Gian-Carlo Pascutto]

But the average branching factor is higher in Go, correct? Around 200 instead of 35 or so I've heard. So it should be easier to parallelize a Go tree.

This is only true if you would not prune anything (minimax). See previous post.