How Intel Validates Their CPUs
By Cade Metx, PC
Magazine, October 20, 1998
Trends Section, pg 28
The Big Test:
Before a CPU sees daylight, it undergoes years of tests.
Try to imagine building a scale model of Manhattan that would
fit on the end of your thumb, duplicating every street, bridge,
tunnel, sky scraper, and corner deli. Fifth Avenue would be
several times thinner than a hair from your head.
Then imagine simulating yesterday's New York traffic on your
model, guiding thousands of virtually invisible cars, trucks,
trains, buses, and bicycles down the many roadways, across
the many bridges, through the many tunnels, and around the
many pretzel vendors. You mustn't let slowdowns occur, and
everyone should get to where they're going exactly when they're
scheduled to do so.
Now imagine doing a similar simulation for every traffic
scenario that could ever conceivably occur, ensuring that
your model will never produce an accident or even a slowdown.
Validating a microprocessor is more difficult.
Processor manufacturers devote extraordinary effort to validation
- in which each new CPU architecture is tested for design,
manufacturing, and compatibility flaws. "Today, all the
major chip companies have their own [validation] test suites,"
says Linley Gwennap, editorial director of Microprocessor
Report. "Intel's is likely the most complex - they've
been at it longer and have more money to put into it - but
the others have extensive suites as well." As a group
of PC Magazine editors learned on a recent visit to Oregon,
Intel spends hundreds of millions of dollars a year on validation,
testing each chip from its conception up to its last days
of production. Much like the preposterous task of simulating
traffic on a microscopic model of Manhattan, it's a miniature
task of epic proportions.
Consider Intel's Pentium Pro, a processor that consists of
nearly 5.5 million transistors. If you decided to validate
each and every one of the chip's potential states, you'd have
to examine 2^5.5 million combinations. That's a 1 followed
by 1.6 million 0s, and it doesn't even begin to explain the
scope of validation. A microprocessor, after all, must smoothly
interact with its chip set, an L2 cache, a memory subsystem,
thousands of hardware peripherals, hundreds of thousands of
software packages, and millions of exacting customers. "Wrap
all that together," says Gerald Budelman, senior manager
of server validation engineering, "and validation is
approximately impossible."
You can't validate a processor by systematically checking
every potential state. Today's manufacturers will never ship
a processor that's errata - free, but they can - with months
of continuous, specialized, and extremely expensive testing
- deliver a working, reliable product.
Even before a Intel chip reaches silicon, it's modeled in
software and tested for months, around the clock and across
the globe. Initially, processor functions are simulated at
remarkably slow speeds from 2 to 5 Hz. Whereas a 400-MHz chip
fits 400,000,000 clock cycles into every second, its software
test model manages 5 at the most. "Windows would not
boot in you lifetime," says Budelman. But at 5 Hz, it's
easier to spot potential bugs and determine their causes.
Subsequently, Intel may model the chip with macroscopic electronics,
creating a literally room-size processor that connects to
a standard PC platform and runs standard software. Built at
a cost of roughly $8 million, the macroscopic Pentium Pro
model clocked at 150 KHz and ran most major operating systems.
Once the chip is rendered in silicon, it's validated on dozens
of so-called Diamondback systems - souped-up PCs with $150,000
price tags. Typically, each pair of Diamondbacks shares over
$200,000 in instrumentation, including full logic analyzers,
PCI monitoring tools, and temperature controllers.
Before the chips goes into production, compatibility tests
are run with hundreds of hardware peripherals and thousands
of software applications. Even after a chip reaches production,
validation continues. Each new chip is tested while still
on the wafer and after it's packaged.
This entire process takes nearly two years, yet it still
fails to find every bug. "In addition to becoming more
powerful over the past few years," says Gwennap, "processors
have taken on more and more complex functions: the floating
point, the primary cache, often a secondary cache. There's
more and more potential for test suites to miss something."
More than 50 errata - all unlikely to affect everyday PC operation
- have been found in the Pentium II since it shipped to customers.
This is not to say that today's validation techniques are
deficient. On the contrary: It's a great tribute to today's
manufacturers that their processors operate as well as they
do.
Return
to Articles Home Page
Return to the FreeHelp
Headquarters
|