[Beowulf] K Computer built for speed, not use

[Prentice Bisbal]

On 10/11/2012 12:37 AM, Mark Hahn wrote:
>> Any general purpose system will inevitably underperform for some people,
>> and many might argue that the art of managing such a project is making sure
>> everyone squawks equally loud about how the stake is being driven into their
>> heart.
> I think of it from the other direction: a specialized machine would need
> to demonstrate really significant savings.

This is the logic behind DE Shaw's Anton computer for molecular 
dynamics, except instead of saving money, the item being saved is time. 
In their papers they argue that designing the unique processors for the 
Anton put them 5 years ahead of waiting for general-purpose processors 
to achieve the same performance. Of course, money was not a concern.

This whole discussion of the K computers practicality reminds me of 
autoracing. The saying is "Win on Sunday, sell on Monday", with the 
logic that winning on Sunday will boost car sales for the "brand" that 
one on the following Monday. The other argument is that the racing 
competition spurs engineering advances that trickles down to passenger 
cars. And we all know that cars that race on Sunday, whether Indy Car, 
Formula 1, NASCAR, or whatever, don't make good passenger cars or 
pickups, and are actually illegal to use on streets.

IMHO, the only computer companies doing the "Win on Sunday, sell on 
Monday" thing right are the companies like Intel and IBM, and Mellanox. 
These companies have many products in the top 500, but are also 
available to the "general public" (in quotes because I'm not sure you 
can call cluster purchasers the general public), and make enough money 
to fund their own R&D.

While supercomputers like the K are awesome feats of engineering, if 
those innovations don't trickle down into products with more widespread 
availability, I have to wonder what's in it for the company and the rest 
of the world (other than government grants keeping them afloat, which 
isn't a great business model in the end)?

I hope I didn't go too far off-topic just then.

--
Prentice