[squeak-dev] Prepare for Thousands of Cores --- oh my Chip - it's full of cores!

Jason Johnson jason.johnson.081 at gmail.com
Wed Jul 9 09:24:41 UTC 2008


On Tue, Jul 8, 2008 at 7:24 AM, K. K. Subramaniam <subbukk at gmail.com> wrote:
> On Tuesday 08 Jul 2008 3:26:14 am Igor Stasenko wrote:
>> It is good to see that they finally trying to use GPUs as generic
>> purpose processing units. But it has own limitations, like inability
>> to run different code in its 2^n working units. So it is more press
>> release than real breakthrough :)
> The discussions on this thread are on the lines of theatre artists discussing
> how to get a movie camera to record stage plays. It ignores a whole new class
> of story-telling possibilities opened up by a such a technology.
>
> A 1000-core commodity computing board is not going to be like a quad-core
> board with 1000 cores instead of 4. It will probably be a hypercube network
> of communicating cores. Nodes are not only "local" or "remote" but can also
> be 'neighbors'. Imagine how you could implement Kedama natively on such
> boards. Instead of having a single-thread VM simulating a world of particles,
> you could create a particle world super VM which would simulate a single-core
> VM as one of its "objects". Keeping a single time references across nodes is
> now easy. This could simplify support for temporal operators.
>
> Imagine!
> Subbu

This is a good point.  I can't imagine the Intel way with the snoopy
bus, write-through cache and so on scaling even to the hundreds, much
less the thousands.  AMD's NUMA is closer, but even that may be a
problem when cores number in the thousands.



More information about the Squeak-dev mailing list