Intel Corp (INTC)

[kpf]

It's a lot, but eight channels would be needed to get close to parity, twelve to exceed it.
I look at Power6 memory architecture as targeted at large database-applications (large off-die L3, high-bandwidth i/o) while Tukwila as designed for its application-base of scientific/HPC apps where latency is paradigm. (Large on-die L2-caches). Insofar i am reluctant to compare the two arches at all, let alone parts of the arches alone.

I raised the question of Tukwilas CSI implementation mainly because i thought it could be a design which appears in Nehalem as well (Which i still believe will be so). Nehalem and K10 will play in the same league, so there comparison is due.

K.