Intel takes inside track to faster chips

Intel commands nearly 80 percent of the worldwide CPU market, according to MicroDesign
Resources Inc. of Sebastopol, Calif. Intel’s 21-year-old x86 architecture is the
standard for almost all PCs and workgroup servers. Last this month, Intel will introduce
new chips to extend the life of that architecture.

The Pentium III, formerly code-named Katmai, will be the newest member of Intel’s
P6 family, which includes the Pentium Pro, Pentium II, Celeron and Xeon. The Pentium III
has the same core design but adds 70 new instructions and eight new registers. This is the
most significant extension to the x86 architecture since Intel stretched it from 16 bits
to 32 bits back in 1985.

Compared with a Pentium II running at the same clock speed—say, 450 MHz—the
Pentium III can execute some 3-D graphical and multimedia operations twice as fast.
Although other types of operations will run at about the same speed, a 500-MHz version of
the Pentium III, also appearing this month, will outrun all current Intel CPUs.

The Pentium III’s biggest impact will come in consumer software, but business and
technical users will benefit, too. The chip’s extensions will accelerate a wide range
of data-intensive tasks, such as computer-aided design and speech recognition.

Around the middle of this year, Intel is expected to introduce a 533-MHz version of the
Pentium III with a 133-MHz system bus. That would be a 33 percent improvement over the
100-MHz system buses of today’s best x86 chips. A faster system bus is particularly
important for servers because it handles the flow of all data to and from the CPU, and
servers typically process more data than desktop PCs.

As for Intel’s rivals, the first seventh-generation x86 CPU is scheduled to appear
in June from Advanced Micro Devices Inc. of Sunnyvale, Calif. The AMD K7 will challenge
Intel’s performance leadership via server-friendly features: a 200-MHz system bus,
better support for multiprocessing and external Level 2 cache as large as 8M.

Cyrix Corp. of Richardson, Texas, a subsidiary of National Semiconductor Corp. of
Sunnyvale, Calif., will introduce an innovative x86 chip, the M3. It will be the first x86
with an onboard 3-D graphics engine. Its other integrated features will likely further
drive down desktop PC prices.

Like AMD, Cyrix wants to widen the market for sub-$1,000 and even sub-$500 home
computers. But highly integrated chips such as the M3 are suitable for business
applications.

Vendors of non-x86 RISC processors are vying with Intel, too. Compaq Computer
Corp.’s Tandem subsidiary is switching its servers from Mips Technologies’
Rx000-series CPUs to the Alpha, which Compaq acquired with Digital Equipment Corp.

Compaq has predicted the Alpha 21264 will hit a clock frequency of 1 gigahertz by
year’s end, months ahead of Intel. IBM Corp. also is aiming at the 1-GHz barrier with
new PowerPC chips.

Overshadowing all these developments in the eyes of long-term planners is Intel’s
determination to introduce a wholly new IA-64 CPU architecture in mid-2000 to supplant the
1970s-era x86 architecture.

The 64-bit IA-64 is a joint development effort of Intel and Hewlett-Packard Co., which
eventually will phase out its PA-RISC processor line in favor of the new architecture.

To achieve faster performance, IA-64 will exploit cutting-edge but heretofore
commercially unsuccessful CPU technology. Its chance of success is iffy. Intel has hedged
its bets by continuing to develop the x86. To ease the worries of managers who have
invested millions of dollars in x86 PCs and servers, Intel is divulging more details about
its plans than ever before.

The first IA-64 processor, code-named Merced, has had a long gestation. Originally
scheduled for 1997 release, Merced is now supposed to appear in mid-2000. Price estimates
run as high as $5,000, almost twice as much as for any previous Intel CPU. The Merced will
go into high-end workstations and servers, not desktop PCs or workgroup servers. Systems
could cost at least $10,000.

In the second half of 2001, Intel plans to bring out its second IA-64 processor,
code-named McKinley, which is supposed to have so many improvements that it will make
Merced seem like a rather crude first try.

McKinley will execute more program instructions in parallel, have larger on-chip cache,
operate at clock rates of at least 1 GHz and run IA-64 software at least twice as fast as
Merced. It, too, will be expensive and will power high-end workstations and servers.

Next in the IA-64 series will be a CPU code-named Madison that should debut in 2002,
courtesy of a new 0.13-micron chip fabrication process that will boost clock speeds to
about 1.6 GHz. Smaller fabrication processes squeeze more transistors onto a silicon die
and deliver higher clock speeds. Intel will make both Merced and McKinley by the larger
0.18-micron process.

Around the same time, Intel will roll out a potentially more important IA-64 processor,
code-named Deerfield—the first IA-64 chip for PCs and smaller servers now ruled by
the x86 design. Deerfield could pull the prices of IA-64 systems down to $2,000 or $3,000,
or whatever is considered reasonable for a powerful PC in 2002. The chip could end the
x86’s long reign as the industry standard.

In the meantime, the x86 won’t be standing still. Within a few months after the
Pentium III’s introduction, Intel will follow up with Tanner, a Pentium III Xeon
variant packaged in a larger, single-edge-contact cartridge for Intel’s Slot 2
interface. Tanner is for workstations and servers, and it will have full-speed, external
Level 2 cache ranging from 512K to 2M.

The Pentium III’s external L2 cache, in contrast, runs at only half the speed of
the CPU core. Full-speed cache is better for servers that handle lots of transactions.

Later this year, Intel plans to ship Cascades, another Slot 2 processor based on the
Pentium III core. Cascades’ 0.18-micron process should let core frequencies exceed
700 MHz; current P6 chips have a larger, slower 0.25-micron process. Cascades will have
on-chip L2 cache instead of external L2 cache to improve performance and reduce
manufacturing cost.

Intel will use the new 0.18-micron process to improve the Pentium III as well. A
version of the Pentium III known as Coppermine, scheduled for release this fall, will
probably run at clock speeds as high as 667 MHz this year and 733 MHz next year.

A higher-end CPU based on the Willamette core, code-named Foster, will likely appear
around the same time as Willamette, or perhaps early in 2001. Designed for workstations
and servers, Foster will have on-chip L2 cache and a system interface four times faster
than today’s Slot 2 processors.

Will Foster be Intel’s last x86 generation? That depends on how IA-64 fares. The
Willamette core should keep the x86 going for a few more years. By then, we will know the
fate of IA-64.

Even if Intel phases out the x86, however, AMD and Cyrix could keep it alive. Other
1970s-era CPUs such as the Z80 and 6502 still survive as embedded processors and
microcontrollers, so x86 is likely to be around for decades.

To ease the transition from x86 to IA-64, future IA-64 processors will be compatible
with software for x86 and Hewlett-Packard’s PA-RISC architectures. Intel recently
revealed some details about how x86 compatibility will work at the CPU level.

An internal decoder will translate x86 programs into IA-64 instructions on the fly. But
Merced will not run x86 software as fast as the best x86 chips available at the same time.
For best performance, developers must recompile their programs for IA-64.

IA-64’s built-in x86 emulation gives some people the jitters. Will it really work?
A good analogy could be Apple Computer Inc.’s transition from 680x0-series chips to
Power PC chips in 1994.

Critics predicted all kinds of problems with the Power Macintosh’s 680x0
emulation, but it worked amazingly well. Before long, Power Macs were emulating 680x0
programs faster than the best 680x0 Macs could run them natively. Although x86 emulation
is more difficult, there’s no reason Intel can’t make it work just as well.

The biggest gamble for Intel is not x86 emulation, it is the outlook for IA-64 as a new
architecture. Intel claims to have designed IA-64 to establish a new industry standard
that will rule the market until 2025. But attempts by other companies to use similar CPU
technology have not succeeded.

Intel is betting that its new twist, reinforced by an already dominant market position,
will make the crucial difference. 

Tom R. Halfhill is a free-lance computer journalist in Burlingame, Calif.