[Info-vax] OT: Intel says 1,000-core processor is possible

[John Wallace]

On Nov 30, 9:25 pm, Neil Rieck <n.ri... at sympatico.ca> wrote:
> Intel says 1,000-core processor is possible
>
> http://www.itworldcanada.com/news/intel-says-1000-core-processor-is-p...
>
> NSR

Possible and useful are worlds apart here.

Eliminating cache coherency and using message passing instead sounds
on the surface a bit like someone re-inventing the Transputer; I
wonder whether there are any relevant patents still in force. TCP/IP
for the inter-core link? I suppose it has a well known API, often well
known for too many buffer copies...

There are very very few problem domains where massive parallelisation
turns out to actually be useful, outside 'the cloud'. In the case of
Intel's proposed chip, it had better be a cache-friendly problem
because keeping those 1000 cores fed with instructions and data is
going to need a lot of bandwidth. If it's not cache friendly that
bandwidth is going to be massively limited by off-chip connectivity,
as many readers here will realise. I can't quite see how 'the cloud'
becomes 'cache friendly', but I presume I'm missing something.

Still, it keeps them looking as though they're doing something
constructive.