Intel Investors! Nice piece on Itanium from MicroProcessor Watch:
*** Editorial: So Long Alpha
By Steve Leibson {7/23/01-02}
By now you should know that Compaq has pulled the plug on the
Alpha RISC processor. Alpha's wheels have yet to spin down, but
the power's definitely switched off. Compaq announced in June that
it will curtail Alpha processor development after 2003 and Alpha-
based systems development after 2004. The company says it will add
two speed upgrades to the existing EV68/21264 processor, finish
developing the next architectural enhancement (the EV7/21364), and
that's it. Alpha software teams will immediately start to target
Itanium, not Alpha. By the year 2004, bye bye Alpha. Also, bye bye
MIPS. Compaq currently uses MIPS processors in its NonStop
Himalaya servers, which the company got through its Tandem
acquisition. Compaq's future servers will be based on Intel's
Itanium.
There can be little debate that adopting Itanium is something
Compaq must do. Through acquisitions, Compaq has put itself into
exactly the same bind that DEC was in a decade ago. Back then, DEC
rode the dual horses of the MIPS and proprietary (VAX) processor
architectures. DEC let neither win, adopting Alpha instead of
either incumbent. By buying DEC and Tandem, Compaq's server groups
now ride three horses: Alpha, MIPS, and Pentium Xeon. Now you know
which horses Compaq shot. But you should have known it all along.
Back in 1996, then editor in chief Linley Gwennap predicted that
workstation and server vendors SGI and DEC would ultimately
abandon their proprietary RISC designs and jump on the Itanium
bandwagon, several years before the wagon got its wheels.
Alpha leaves a significant legacy in the history of
microprocessors. It was designed like a racing car: speed first
and foremost. Throughout the 1990s, Alpha was the speed leader in
integer performance by a wide margin, thanks to its combination of
a simple (hence fast) instruction set, dynamic logic, and
bleeding-edge circuit design. Circuit wizards in the Alpha dungeon
pushed DEC's relatively low-volume manufacturing facility to
produce faster parts at a given lithography level than commercial
semiconductor vendors could hope to achieve. Only occasionally did
competitors catch DEC, when the company was slow to make a process
upgrade.
The Alpha processor joins a number of other worthy microprocessors
on the historic scrap heap. There's the AMD 29000, a RISC
processor originally designed to be a Unix workstation engine,
which became a successful embedded processor before fading out.
There's Intel's 960 series, with similar design heritage and a
similar fate, after bloodily wresting the laser printer market
from Motorola's 68000. There's Motorola's innovative 88000 family,
which could never crawl out from under the shadow of the 68000 and
that possibly put the stake into Data General. Ditto the ill-fated
Fairchild Clipper, which inflicted severe collateral damage on
Intergraph before Intergraph cast off the demon. Both Data General
and Intergraph dumped their essentially proprietary RISC
processors for Intel's P6. The list of fine-but-obsolete
processors continues to grow.
From a technical standpoint, it seems to me that Itanium, or
something like Itanium, has to win the microprocessor wars for
high-end servers, because VLIW (or EPIC in the case of Itanium)
simply makes more sense than continuing the increasingly difficult
battle to augment RISC processors' superscalar capabilities. One
of the original advantages of RISC design, reduced transistor
count, has long since disappeared with the addition of superscalar
operations as a speed enhancer. The additional layering of
superscalar bells and whistles increases a processor's IPC rating,
but with a quick limit. Shooting for more than four-issue
superscalar operation is a dead end.
Conversely, the VLIW approach relegates the search for code
parallelism to the compiler, thus allocating one or more
compilation processors and relatively abundant compile time to
this task, in stark contrast to the relatively few transistors and
real-time constraints that superscalar approaches impose. Argue
all you want about the elegance of the Itanium architecture, VLIW
is just a better approach to instruction-level parallelism than
superscalar RISC is. As a high-end RISC race car, Alpha is dying a
natural death.
On the business side, Compaq's decision also makes a lot of sense.
Compaq is a company built upon, and steeped in, compatibility. The
company was founded on the concept of close IBM PC compatibility
in a (purportedly) portable form. Taking an industry-incompatible
approach to high-end servers seems to be an unfortunate exercise
in corporate dissonance. It's good to see Compaq's leadership step
up and make a hard decision to end the experiment.
So who remains? Sun and IBM. IBM seems unlikely to wind the entire
company around the Itanium pole. Unlike Compaq, IBM prefers to
create standards; it doesn't follow them without a strong
strategic reason. The company has successfully built systems to
its own standards for decades. IBM seems well committed to the
PowerPC from one end to the other. Yet if you look at Intel's Web
site, you'll find IBM listed as an early Itanium adopter in both
the server and workstation categories. The camel's nose is already
under IBM's server tent.
I think it will be a very long time before Sun Microsystems
accepts an Itanium-clad future. Sun has already set one of the
microprocessor standards in the workstation and server market
(after chucking the 68000 many years ago). Sun has captured nearly
half the server market, so there's little for the company to gain
by adopting Itanium today. In fact, now that Compaq has declared a
concrete end to Alpha systems, you should expect Sun to quickly
try to gain market share by aggressively targeting existing Alpha
system customers as likely sites for conversion to SPARC-based
servers. After all, if a Compaq server customer knows it must port
its software from Alpha, it's just as easy to port to SPARC as to
Itanium, and it's probably safer too, because the compilers and
associated software-development tools for SPARC are more mature
than those available for Itanium. Sun will need to convert to
Itanium only when Itanium-based servers and workstations start to
consume some of Sun's market share. That won't happen this year. |
