> Multi level cache, translation lookaside buffers and a number of other items
come to mind.

Segment Table Origin Register Save Stacks.  Each TLB entry is "tagged" with the STO it
corresponds to.  If an address space switch occurs, a new STO is loaded but the old 
one is
saved in a stack and its TLB entries are tagged.  If the old one is ever restored, the
corresponding TLB entries are immediately returned to validity.  Came in with the 
3033, I
believe.

And other stuff.  Loads of it.  Cycle time isn't even a valid comparative metric 
between
different generations of IBM processors, except that it shows what technology level 
they're
using.  And I'm not sure that is very interesting.

Different CPU designers used different degrees of this stuff at different levels.  In 
the
1980s, it seemed that IBM's designers were aiming at a sweet spot around 70% CPU - 
Amdahl's
aimed at 95%.  The relative performance of these systems changed with CPU load - the 
Amdahl
machine got "faster" the more it was loaded.

--
  Phil Payne
  http://www.isham-research.com
  +44 7785 302 803

Reply via email to