Jeff K (jeffk.delete@this.smthng.awfl.com) on November 22, 2013 7:17 am wrote:
> I suppose a new custom core is the best guess, but is it possible to have some kind of hybrid of cores?
> Like, say, clusters of 8 KNC cores for throughput + 1 Core core for the heavy lifting + a LLC slice?
> Impossible to tune the softwares to fit something like that? Or just a bad idea to begin with?

Judging by the raging success that was Cell, I'd say it's both "impossible to tune the software" AND "just a bad idea to begin with".

Of course as with everything there are counter-arguments:

1. CPU + GPU systems effectively behave as you describe today. In a prior job I architected the datapath for a very large/fast (multi-million dollar, many billions of pixels/sec) imaging system, and I partitioned the work between Xeons and GPUs basically as you describe.

2. In addition to having a heterogeneous architecture, Cell went out of its way to make development for the SPEs difficult. You could therefore reasonably argue that its problems were as much in the implementation as in the context.