Re: Chipsandcheese article on the CDC6600

Michael S <already5chosen@yahoo.com> writes:

At the end, the influence of 6600 on computers we use today is close to
zero. On the other hand, influence of S/360 Model 85 is massive and
influence of S/360 Model 91 is significant, although far less than the
credit it is often given in popular articles.

Yes, all modern computers have virtual memory (which started with
Atlas (and later S/360 Model 67), they have caches, which started with
Titan (and later S/360 Model 85), they have reservation stations
(which started with S/360 Model 91).

However, the main reason why reservation stations won is because
hardware branch prediction outpaced compiler branch prediction since
the early 1990s*, and because the reorder buffer was invented, neither
of which is due to anything done in any S/360 model or the CDC 6600).

If hardware branch prediction had never been invented or had turned
out to be a dud, maybe we would all be using EPIC architectures that
use scoreboards rather then reservation stations; or maybe the
register interlocks that were used in advanced in-order RISCs (those
that Mitch Alsup calls OoO) and AFAIK in IA-64 implementations were
good enough and one would have done without scoreboard.

[*] More supercomputing-oriented people may claim that it has to do
with the number of in-flight memory accesses, but actually IA-64 shone
on SPEC FP (where in-flight memory accesses are more important than
for SPECint), so it seems that there are ways to get the needed
in-flight memory accesses with in-order execution.

Back at their time 6600 was huge success and both Model 85 and Model 91
were probably considered failures.

From what I read, the Model 91 was considered a technical (and
marketing) success, but commercially a failure (sold at a loss, and
therefore quickly canceled).  But apparently the market benefit was
enough that they then built the 360/195 and 370/195.  15 91s were
built and about 20 195s.  The 195 was withdrawn in 1977, and AFAIK
that was the end of IBM's supercomputing ambitions for a while.  This
may have had to do with the introduction of the Cray-1 in 1976 or the
IBM 3033 in 1977.  IBM eventually announced the optional vector
facility for the 3090 in 1985.  OoO processing vanished from S/360
successors with the 195 and only reappeared quite a while after it had
appeared in Intel and RISC CPUs.

The Model 85 was built only 30 times, but it was the fastest IBM
machine until the Model 195 (the CDC 7600 was faster, though, and also
a bit faster than the 195).  And the cache was then included in the
Model 195, and on the 303x, and I expect all later IBM mainframes.  So
that certainly was a success.

Concerning the CDC 6600, the barrel processor features of its PP can
be considered predecessors of modern SMT (about as close as Model 91
reservation stations are to modern OoO). 

Concerning the Atlas, it's interesting that a project intended as a
supercomputer introduced virtual memory, while Cray rejected VM for as
long as he designed the CPUs.

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