Re: 1teraflops cell processor possible?

In article <[EMAIL PROTECTED]>,
Glew News SBC <[EMAIL PROTECTED]> wrote:
>"Del Cecchi" <[EMAIL PROTECTED]>
>> In just rereading the IBM J. of Research and Development article
>> http://www.research.ibm.com/journal/sj/402/allen.html
>>
>> I was struck by the following paragraph.
>> --------------------------------------------------
>> More generally, thinking on algorithm and system design is still dominated
>> by the invalid perception that compute units are the most expensive
>resource
>> in a computing system. Applications are designed to minimize the number of
>> operations required to solve a problem. Systems are designed with enough
>> memory and I/O to achieve a high usage of the compute units, even if this
>> leads to low memory or disk utilization. In fact, the cost of systems is
>> dominated by storage cost, and by the cost of the logic required to move
>> data around (caches and buses). If one thinks anew from basic principles,
>> the conclusion will often be that algorithms that use more computing but
>> less memory, or use more computing but require less communication, coupled
>> with systems designed to ensure more effective use of memory and
>> communication, even at the expense of lower utilization of the processing
>> units, are more cost-effective than conventional algorithms and systems.
>>
>> This is from 1999.  Is this the paper referred to earlier?
>
>Amen Brother Del! Or Brother Allen per Brother Dell.
>
>The best way to design computers nowadays is to start
>off from the memory subsystem, figure out the best cache
>(or streaming, Brother Robert) hierarchy you can build,
>match the register file, bypass network, and scheduler,
>and only lastly to figure out how many ALUs and FMACs
>are needed to consume everything that you can feed them.

Well, yes, but ....

Back in the mid-1970s, that was heresy - though it was actually first
predicted in the 1960s!  But everyone with Half a Clue has known it
since the 1980s - though that doesn't dispute what the abstract says,
given the proportion of the Clueless :-(

However, the abstract contains one error of fact - storage cost is
NOT always the dominating factor, though I can see why an IBM article
would say that it is.  It isn't rare for communication latency to be
the dominating factor in cost, and I am not just talking about HPC.

Also, I think that you aren't being radical enough.  I would start
off with the memory access MODEL, because it is clear that the current
one badly needs rethinking.  And this could easily require application
redesign, as that abstract says.


Regards,
Nick Maclaren.