For whatever it
is worth, this was a discussion about compatibility. My point
was and is, Intel owns the trademark; and defines / continues
to extend the INTEL*64 and IA-32 ISA's. The current
definition can be found at:
Intel has
invested heavily in the ability to moved customer codes from
4004 to today's Phi ISA and IMO, done an excellent job of it.
Certainly from the 386 family and later. As Johnny and I
both said, you can run MS-DOS and old DOS programs on my
current system from an Edison (IoT) module that costs a few
dollars all the way up to a world largest supercomputer (the
Milky Way 2 system in China) - which is a bit of a frightening
thought to me.
As Tim points
out, the cost to do that compatibility is huge on the
development / investment side. As I said, the on-die tax has
been reported to me my my brethren as about 5 %.
So coming back
to compatibility, when the first processors to support
INTEL*64 were created, Intel's engineering team had a choice.
What is interesting is that Intel's engineers chose to
ensure that current set of applications codes continued to
run. They did not have too. We might have had two
completely different instruction sets if they had chosen
otherwise. I personally think, we as consumers won because of
that.
IMO: I think
that would have been a bad thing for developers because a
number of choices would need to be made and confusion would
have likely arisen. That said, if you look at Apple's use of
Fat Binaries, it might have been manageable if Linux and
Windows had done something like that and the different
compilers generated code that way. As it turns out, that is
not so far fetched, since we are doing that today for Phi, but
the difference is that Intel owns both definitions and builds
a single set of development tools so that the differences to
user is unseen.
Frankly, I'm
personally glad that the folks that made decision took the
path of starting with an existing set of extensions. But
that was 10-15 years ago. Intel has continued to extend the
ISA since that time and I believe that we will see that going
on into the future. As I said, we already see more extensions
with the Phi product line. But back to my point, a pure
core binary programs will run on a Phi, although today, Phi
programs will need to be emulated on Core (until such time as
the Phi features and extensions make it into the primary ISA).
Will that happen? I can not say, but given the history of
Intel, I would suspect it might because compatibility has been
something Intel has heavily invested.
I probably
should add in all of these comments, they are my own and not
necessarily those of my employer. That said, I openly point
out that I'm a Sr. Architect in the HPC team @ Intel.
Clem
