[rescue] Sun Kit Needed for EE Student Here

[nate at portents.com]

> I wonder whether part of the speedup is due to faster memory bandwidth
> in the Opterons.

Not just memory bandwidth, but memory latency is also key.  If you compare
the performance of the G5 vs. the AMD64 architecture, the AMD64 will
perform better in floating point when the data doesn't fit in the cache
size of the G5 because the memory latency is so low.  When the data does
fit entirely in the G5 cache however, in crushes the AMD64.

It's also why the new socket AM2 AMD64 chips with their DDR2 memory
controllers (twice the bandwidth, but higher latency because it's DDR2)
have only a 5-10% performance increase at the same clock as their older
DDR AMD counterparts - low memory latency is a big reason why AMD64
performs as well as it does.

It's interesting to note the POWER architecture approach to solving memory
latency by having up to 128MB of L3 cache..

- Nate