Alberto (git.delete@this.git.it) on August 14, 2018 12:57 am wrote:
> Likely because there are known limits at the amount of heat that a semi-passive cooling solution can dissipate
> from a single socket. Dividing the heat on two or more sockets the OEMs have an more easy life to think about
> an efficient semi-passive cooling via copper heatsink under a moderate air flux inside a tiny server box.
>
> You can follow IBM with some exotic solutions but this is not absolutely the case. Over 200W/socket the
> heat dissipation become very serious to manage in actual popular server boxes. Finer the process utilized
> the thing become even worse cause the stellar power density on die, and Fins do not help at all.
> My bet the shrink game will be stopped mainly by power density rather than by yields issues.

In theory, the chiplet approach is supposed to address both yield and power issues (in cases where single-threaded perf is less important at least), just by being able to bin out the best execution units and assembling them together and making substantially larger aggregate parts running closer to their perf/Watt sweet spots. As the Anandtech review pounds on though, interconnect fabrics need more refinement in order to not overwhelm these gains with overhead, so maybe we'll need more EMIB/active interposer mesh solutions in the future instead of the passive organic substrate networks we see in this generation of Zen.