[rescue] i860 Success

[Charles Shannon Hendrix]

Sat, 01 May 2004 @ 15:25 -0700, Francisco Javier Mesa-Martinez said:

> On Thu, 29 Apr 2004, Charles Shannon Hendrix wrote:
> 
> >
> > The i960 was a re-design of the IA432, fixing the speed issues and was a
> > decent general purpose CPU with an MMU and everything.
> >
> > Intel *marketing* "designed" it as an embedded CPU by removing the MMU
> > and marketing it as an embedded processor.
> 
> Whaaaa? I never heard that, the 432 is one (if not the most) CISCiest
> projects ever attempted, tons of Opcodes and HLL hardware support in
> silicon. The 960 is pretty RISCy, I don't think it ever had an MMU, but I

Yes it was best called VCISC, but the i960 is hardly RISC in the
traditional sense.

The i960 came out of the Gemini/P7 project at Intel, which was based on
the 432 design, but with changes to the ISA to lean toward RISC.  It
still had a lot of the design and ideas in there, including the OO parts
and tag bit.

The goal was to have the feature set of the 432, but speed could beat
something like a Motorola 68K.

I think it was the P7 that was used in BIIN and also some kind of Air
Force on-board computer, not the i960 as I agreed to in another message.

After that, Intel removed the OO parts and the tag bit, and that became
the i960.  The i960 was a very good CPU, but Intel marketing buried it
because it would have competed with the 80386 which was in competition
with it at Intel.  Intel actually finished the i960 first, but many
people claim they sat on it while the 80386 was finished.

Intel had a golden opportunity to get away from x86 right there.

> the 432, there is not much in common... except that the 960 is almost
> like an anti-432, they dropped everything that made the 432 like the
> complex ISA, the microcode, the tag support etc.

They didn't drop everything though.  

> It was a fairly cheap, clean simple design and I assume it made it as
> an embedded design because of those three factors. Plus intel did not
> want anything to interfere with the x86,

It was not designed as an embedded CPU, that was a marketing decision.

> I believe the 960 was also developed as a contract with the DoD for
> embedded processors.

No, that was the 432.  It was developed after the DoD "everything in
Ada" edict.

http://www.brouhaha.com/~eric/retrocomputing/intel/iapx432/

The P7 was used in a DoD project, but was not designed for it.

These days, we have the transistor budget to make chipset like the 432
fast enough for most any task.

> The i860 was a completely different machine, it was supposed to be
> the main competitor of the 88K series from motorola. Both chips were
> introduced in '88 I believe and were supposed to take over from the
> early RISC adopters.

Well, I saw the Intel road show in two settings: academia, and an
embedded shop.  I'm sure it colors my perceptions.  I should have
mentioned that in the first place.

After thinking about it a bit, I think maybe they pitched it as an
alternative to x86 because almost everything we did in college and in my
first embedded work was Motorola.  In fact, when I did see Intel or even
Z80, it was usually when we were replacing it with Motorola hardware.

Every shop we visited and most of the industry rags I read were all
about Motorola.  It was especially true if you did your work on VMEbus.

The Intel marketing droid may have been saying, "If you find x86 a bad
fit, we have the 860 which is much better..."

> However the 860 was hard to code for because it had some VLIW
> characteristics,

...among other things.  It also let you see the pipeline from userland,
though that was a hard feature to actually use.

> think Alliant also made a few minis based on this chip. But it was hard to
> get anywhere near the quoted performance figures so it was never a popular
> design.

The only way to get good performance was with a relatively small set of
code which was very well scheduled, and preferrably fairly small.  It
was great as an embedded controller or GPU.

You also wanted to avoid interrupts as much as possible.  If you were a
wizard, I think you could order your code to make them less an issue.

> Think of it as a modern day Itanium, a pain in the neck to code for, and
> even harder to get the performance Itel quotes in their benchmarks.

That was my first thought when Itanic hit the street with a wet squish.

I still like Alpha better, even though it is old and crusty now.

-- 
shannon "AT" widomaker.com -- ["Tara is grass, and behold how Troy lieth
low--And even the English, perchance their hour will come!"]