Re: Larrabee details: Yes, it is based on the Pentium. :-)

John Dallman <jgd@xxxxxxxxx> wrote:
It appears that Intel's compiler people scored many brownie points
during the architecture design by saying "we can deal with that" about
many of the sticky problems. Sadly, my source for this is covered by
NDA. I have also theorised that both Intel and HP hoped that they could
get the other company to do the hard bits.

It certainly seemed, even at the time, that HP's management were great
believers in silver bullets for software problems. A few years earlier
HP's then-president John Young had commented that though software
advances had not kept up with hardware evolution lately, they expected
the software problem would be "largely solved" in the next few years.

In the Tahoe development era HP VPs were making wildly optimistic
comments about PA-RISC->IA-64 dynamic translation performance, and
the general ability of magic compilers to solve performance and
complexity issues.

I think many of us wondered why they seemed to be going to so much
trouble to turn a hardware problem into a software one.

But they had been extremely successful doing all these things in the
previous architectural transition to PA-RISC, so it's possible to
see where some of these beliefs might have come from, but they had
some very clever people working on the PA-RISC transition in its
day and the problem was *much* simpler than the IA-64 transition.

IA-64 represented an exponential increase in complexity but
everyone at HP seemed to believe that the effort would simply
be a linear increase over what they had done the last time. How
they managed to have nobody in a position to question this still
boggles the mind.

Someone here recently complained that one of IA-64's problems was
the abandonment of x86, but actually I think the problem was more
the reverse in that they insisted on hardware binary compatibility
with x86 to the point that (IIRC) a Merced chip would reset into
386 real mode and would probably boot and successfully run DOS or
Windows 3.1 without software changes (at least in the boot process).
I have no idea if, or how much, the x86 compatibility requirements
affected the implementation schedule.

I think the primary failure of Tahoe/IA-64/Itanium was just that
it took too long to implement so that it was not performance
competitive when it came out. Also by then the x86 revolution
had taken place in that x86 went from a crap architecture that
everyone was convinced must go away eventually into one with
ultra state-of-the art implementations and performance (thanks
mostly to competition from AMD). It was thus clear that the
problems IA-64 was designed to solve were not actually the
problems people were going to have going forward. Had IA-64
made it out a few years earlier then this might still have
been true, but I don't think it would have been as obvious and
so the "EPIC" idea might have had a much better chance of
getting established and reaching a critical mass.

I think the best evidence of the failure of the "EPIC" style of
architecture is the consideration of making an out-of-order
version of it.

G.
.

Relevant Pages

  • Re: Hinted conditional branch ?
    ... A marketing name of IA-64. ... Architecture, and Itanium Processor. ... Tahoe where the HP architecture guys sold the Intel ... As x86 became the real 64-bit Intel architecture, ...
    (comp.arch)
  • Re: Interview with Alan Kay
    ... This was similar to the 432 experience, and Intel and Siemens ... Anybody who says that Itantic's competition ... EPIC is a very interesting architecture and is successful technically ... doesn't run x86 code as fast as Xeon and AMD processors. ...
    (comp.lang.lisp)
  • Re: Studio 10 compilers for Solaris released
    ... It's not the Sparc architecture that's the problem, ... * IA64 doesn't run previous binaries well enough to be competitive, ... since converting 32 bit x86 binaries to an engine tuned for VLIW ... the desktop technical computing market ...
    (comp.unix.solaris)
  • Re: x86 architecture concepts
    ... I have better answers - I know a fair amount about the ARM, x86, and PPC ... There are two aspects of a processor architecture - there is the ... means they have the same instruction set and register build-up. ...
    (comp.arch.embedded)
  • Re: Why we build big computers
    ... be that it *isn't* x86. ... many machines of whatever architecture as it wants to build malware, ... there just aren't as many Itanium hackers out there are there are x86 ... Even if a hack exploited some knowledge of the ia32 or AMD64/EM64T ...
    (comp.arch)