Earth Simulator director says cheap innovative modification alllows 16 PETAFLOPS.

[NANOTEC]

the director of the Earth Simulator said in an interview with Nature magazine, published last week, that he was planning an inexpensive and innovative modification to that machine that would give it power equivalent to a 16-petaflop conventional computer.

http://www.iht.com/articles/2005/08/18/business/super.php

The interview in Nature can be found here but you need to subsribe or be a subsriber and it's not free.

http://www.nature.com/nature/journal/v436/n7052/full/436763b.html

What does this mean for the other supercomputers trying to compete for the top position?

 

[Blazkowicz]

how can they cheaply turn a 36 teraflops surpercomputer into a 16 petaflops one?

 

[Pete]

Obviously the other supercomputer teams also need to hire MacGuyver.

 

[nutball]

how can they cheaply turn a 36 teraflops surpercomputer into a 16 petaflops one?

Overclocking 4 teh win!!1!!

 

[NANOTEC]

how can they cheaply turn a 36 teraflops surpercomputer into a 16 petaflops one?

Well you'll need a subsciption to Nature to find out. Of course they will not reveal all the details for their competitiors to copy. Also since it's coming straight from the horses mouth and printed in one of the top scientific journals in the world, I have no reason to believe it's a false claim. In fact if it turns out to be as good as he claims, it would be a major breaktrhough in the field of HPC. Very exciting times ahead.

 

[pc999]

Could Cell be involved?

 

[Fox5]

I'd imagine some sort of cheap DSP add on that's only useful for some very limited calculations.

 

[NANOTEC]

Could Cell be involved?

I doubt it for at least a couple of reasons.

1. CELL is not an NEC diesign. The ES is an all NEC design.
2. CELL's double precision isn't all that great and single precision is pretty useless in scientific computing.

I'd imagine some sort of cheap DSP add on that's only useful for some very limited calculations.

I was thinking about that too, but then I realised that though it's a viable option, it may not fit into the "cheap and innovative" description.

 

[pcchen]

Of course, "equivalent" is probably the keyword (I've not read the interview, so it's just my guess).

The Earth Simulator is a kind of its own, not only because it's a vector machine, but it's massive interconnection. There are more cabinets for interconnection devices than other devices. The inter-node bandwidth is more than 12GB/s bi-directional. This make the Earth Simulator much more efficient than "conventional" supercomputers, probably twice or more.

I believe that they can upgrade the CPU without touching the interconnection, while maintaining similar efficiency. NEC's SX-8 actually provides twice performance as SX-6 (used in the Earth Simulator). However, if the Earth Simulator can be four times as efficient as a conventional supercomputer, it still needs 100 times more computation power than its current form. It's hard to imagine how they could achieve this.