IPC (was: Reservation stations)

EricP <ThatWouldBeTelling@thevillage.com> writes:

MitchAlsup1 wrote:

[...]

(Nobody bats an eye at using kilobytes for the branch predictor.)

IIRC it's in the hundreds of KB on the larger cores.  But it pays
off.

One should also note that these machines deliver 1-2 I/c RMS
regardless of their Fetch-Decode-FU widths.

>
Yes, its mostly all for naught.

Evidence?  If it was all for naught, people would not build these big
cores.

Counterevidence:

http://www.complang.tuwien.ac.at/anton/tmp/opt-ipc-uarch.eps

Goldencove has a median IPC of almost 4 for these benchmarks (a bunch
of Forth benchmarks running on Gforth).  You also see a steady advance
between generations, from Sandybridge to Goldencove, from Silvermont
to Gracemont, from K8 to Zen3, and also the expected differences
between A72, A76, and Firestorm.  And the in-order cores are the
dashed ones at the bottom.  Particularly remarkable is Gracemont,
which usually has better IPC than Skylake and comparable IPC to
Firestorm.

- anton
--
'Anyone trying for "industrial quality" ISA should avoid undefined behavior.'
  Mitch Alsup, <c17fcd89-f024-40e7-a594-88a85ac10d20o@googlegroups.com>