 |
Overview
Intel already had a robust track record with microprocessor development when
IBM* chose the 8088 processor as the heart of the 1981 IBM PC. The 8088 was a
16-bit, third-generation microprocessor that followed "Moore's Law" as shown in
the left of Figure 1. Gordon Moore made his first observation about the "doubling
of transistor density on a manufactured die every year" in 1965, just six years
after he invented the planar transistor and four years after he and Bob Noyce produced the first
planar integrated circuit. Gordon admits that, initially, he did not expect his law
to still be true some 30 years later, but he is now confident that it will be true
for another 20 years as shown in Figure 1.
|
| |
Figure 1: Moore's Law Spans 45 Years.
(Click here for a higher resolution jpeg image.)
|
Economic Challenges
The economic result of Intel's relentless pursuit of Moore's Law is actually more
remarkable. The average price of a transistor has fallen by six orders of magnitude
due to microprocessor development. This is unprecedented in world history; no other
manufactured item has decreased in cost so far, so fast. It is interesting, therefore,
to look into this silicon technology to understand its scope and to be able to predict
its future.
Intel's microprocessors, and many other integrated circuits, are manufactured using
a planar process where a pure silicon wafer is selectively masked and diffused with
chemicals to make multiple transistors. This combination is then selectively masked
again and metal is deposited on the wafer to interconnect these transistors. The
first integrated circuit used only one layer of metal; today's Pentium® II processor
uses five layers of metal to increase the packing density as shown in Figure 2.
|
| |
Figure 2: Cross-Section of Five-Layer Metal.
(Click here for a higher resolution jpeg image.)
|
Transistor density is the key price driver of integrated circuits. The area on a wafer
is a fixed cost (about $1B/acre) so product complexity and price will depend upon how
small the transistors can be made. Making transistors smaller is "all benefits":
smaller transistors operate faster, interconnections are smaller, system reliability is
increased since more functions are integrated into a single place, the power is lower
and the cost is cheaper. Everything gets better and there are no real engineering
tradeoffs to make. The results that Intel has achieved in its microprocessor development
are shown in Figure 3. Notice how the die from each processor gets progressively
smaller with each new silicon process and notice how the next generation processor is
then larger with more transistors on it. The cycle then repeats: smaller, faster, and
cheaper. The market demand for higher performance Intel processors is growing exponentially,
so making them smaller also allows us to meet customer demand.
|
| |
Figure 3: Intel Microprocessor Development.
(Click here for a higher resolution jpeg image.)
|
Intel's challenge in making transistors smaller consumes a large R&D budget and
results in new factories being built and retooled for each new generation of
microprocessor. The manufacture of processors in large volumes is a capital-intensive
business which now costs over $3B/year. There have been dramatic changes in manufacturing
to meet today's demand. In the 1970s the technicians wore smocks and used tweezers to
move the wafers from one process step to the next. A fab plant today is quite different:
The technicians wear self-contained bunny suits (the real ones are white, they turn shades
of metallic when put close to a marketing person!) and they work in an extremely clean
environment (over 100 times cleaner than the best operating theater) with robots moving
large numbers of wafers between the multiple processing steps.
Atomic Challenges
So where do we go from here, and what are the challenges looking forward? Today we have the
ability to integrate a doubling transistor count every 18-24 months. The challenge is
deciding what to integrate, how to manage the increased complexity and how to validate
the design. We must also consider interconnection delays, increasing power dissipation
due to the large transistor count, and the shorter wavelengths of light used in the
photolithography process due to the smallness of the devices. All of these challenges
are solvable with money and effort. One challenge, however, that will be difficult for
even Intel to solve, is the fact that when transistors get smaller the gate thickness
starts to approach the atomic nature of matter! It is postulated that the thinnest
gate cannot be less than ten atoms thick and at our current rate we should reach this by
the year 2017. Not a major concern today, but something that Intel has a research team
working on.
What to Integrate
There are many computer architecture techniques that can be employed inside the processor
core to increase its performance. Fundamental techniques such as cache memory and a
floating point unit were added to the processor core early on. Advanced techniques such
as pipelining and multiple instruction execution were added with the Pentium® processor.
Leading-edge architectural techniques such as speculative execution and data-readiness,
run-time instruction scheduling were added with the Pentium Pro processor. Computer science
is now looking foreward with a changed paradigm: Don't be constrained by the hardware
implications of an approach; we'll have the technology within a short time to implement it.
Transistors within the processor core communicate with each other at extremely high
speed—the Pentium II processor is at 300MHz today and this core frequency will continue
to increase in 1998. When communicating with
transistors on other devices the signaling must go via the relatively slow
component-to-component interconnect. This has the potential of reducing
system performance. A simple solution is to integrate more of the system
hardware functions into the processor core. This must be traded off with the
cost effectiveness of implementing the system function in software or in dedicated
hardware or a combination of the two. This hardware/software balance will be
continually optimized as we move into higher performance, future processor generations.
Managing Complexity
The increase in transistor density has caused a similar increase in engineering staff.
There are now many hundreds of architects and design engineers working at numerous Intel
sites all over the world on next-generation processor implementations. The project
management challenges to keep such a large team focused and productive requires a
sophisticated computer network with advanced workstations and development tools. It is
interesting to note that only the very latest generation of Pentium II processor based
machines has enough horsepower to be used in creating the next generation processor! And
so the cycle continues.
Design Validation
Design validation and verification is now an integrated part of the design process; in the early days, it was done after the processor was designed. Intel has approximately
the same number of engineers testing and validating its processor designs as those
architecting and designing. Additional circuitry is added to the processor core to
ease the debug and validation process. Test coverage models are generated so that
each and every logical function can be thoroughly tested. Additionally, legacy
software tests are run on the computer model to ensure software compatibility with
previous processor generations.
Interconnection Delays
As the geometries of a transistor get smaller, the propagation delays through
the device also get smaller. Unfortunately, as the metalized interconnects get
smaller, their resistance and capacitance increases and therefore the propagation
delay through them increases. As we move to 0.25micron geometries the delay through
the metal is more than the delay through the transistor—not at all what we learned
in engineering school! Intel has multiple approaches to this challenge—the most
obvious being to use a metal with a higher conductivity than the aluminum currently
used in production. Copper offers some attractive solutions but this is more difficult
to process. Alternative design techniques that use more transistors and less
interconnect are also being investigated.
Power Dissipation
As the transistor count goes up then so does the power dissipation. More important,
however, is the corresponding increase in frequency, since power increases as the square
of the frequency. The only variable that is changeable in the power equation is supply
voltage—power dissipation is also proportional to the square of the voltage. So we
must reduce voltage as we move forward into the future. The "same-power" voltage for an
0.18micron process is 0.5 volt, compared with 3.3 volts today. But operating at 0.5
volt will present new challenges, since a few millivolts of supply variation or noise
across the processor core will now be a significant percentage of transistor switching
voltage.
Lithography Limits
Today's 0.35micron process uses visible light in its production lithographic process.
Intel's 0.25micron process has moved to deep ultraviolet. Research into the 0.18micron
process is pushing deeper into the ultraviolet spectrum: the benefits of 193nm light is
that the same optical tools and process methods can be used. Good experimental results
have been obtained in the lab but tools do not currently exist for large area coverage
or for a production environment. Looking forward to 0.13micron, Intel is researching
extreme ultraviolet, x-rays and direct-write electron-beam techniques.
Micro 2012
There are no theoretical or practical challenges that will prevent Moore's Law being
true for another 20 years at least—this is another five generations of processors.
Using Moore's Law to predict into 2012 Intel should have the ability to integrate 1 billion
transistors on to a production die that will be operating at 10GHz. This could result in a
performance of 100,000 MIPS. This is the same increase over the Pentium® II processor
as the Pentium II processor was to the 386! This prediction is so staggering that it
borders on unbelievable—but our increased investment in silicon R&D has continued
to produce these kinds of results. Intel sees no fundamental barriers in its path to
Micro 2012, and the theoretical physical limitations of wafer fabrication technology
won't be reached until the year 2017.
Impact on the PC Platform
For the first time since the original PC, we have processor performance being applied
outside the primary needs of software applications. Some applications will continue
to drive for maximum available performance but others will use the processor capabilities
in other ways. Mobile systems are able to slow the processor
down to conserve battery life and still
deliver incredible performance to the user. PCs targeted at
home entertainment could use
the processor performance to decompress an MPEG-2 video stream in software. This
reduces the cost of the PC by removing specialist decompression hardware making
the system more affordable to a larger audience. The operating system can also take
advantage of this increased processor performance by doing system integrity work such
as virus detection, software upgrades, file consolidation and systems management
as background tasks. The delivered processor performance is allowing the PC platform
to evolve out of its "one-size-fits-all" paradigm into purpose-built systems which
match customer requirements as shown in Figure 4. These systems are all based upon
the same processor core so they all run the same operating system and support the
same applications. Each system is tuned on top of this baseline capability to deliver
features targeted at its respective audience. Applications are designed to be scalable
across multiple platforms, and a "high-end" system will typically deliver a better user
experience.
|
Figure 4: PC Platform Purpose-Built Systems.
|
New PC Platform Usage Models
An area which has taken full advantage of the increasing silicon technology is graphics.
It is now possible to get amazing 2D and eye-popping 3D graphics at volume desktop price
points. This is a natural evolution of the desktop platform and the term "visual computing"
has been coined to describe this recent phenomenon (see last month's Platform Solutions
Focus section on Visual Computing).
It is now just as easy to create a full-color video clip with sounds as it used to
be to create a graphic layout—and if you are trying to sell a vacation in Hawaii,
then it is much more compelling to use a video with waves softly crashing in the
background than to use a static graphic. The Internet is changing how people
interact. When combined with visual computing capabilities, businesses and consumers
will have the ability to interact "screen-to-screen," no matter where they are located.
Lower Cost Peripherals
Another area that processor performance is used to great effect is the lowering
cost of attached peripherals. The modern peripheral is essentially a "sensor-on-a-wire,"
and any required signal processing is done by the host processor. A digital camera,
for example, captures its picture on a CCD sensor and passes this raw data via a
Universal Serial Bus (USB) cable
into the PC platform; the host processor implements color space conversion, aspect
ratio correction, scaling and interpolation, gamma correction and white balance.
The software could also correct for dead-pixels in the sensor and will manage user
preferences. The same technique is being employed on scanners, photo-printers,
plotters and a variety of new gaming peripherals. This reduces the cost of the
platform by removing specialist hardware and thus makes the system more affordable
to a greater market.
Summary
The silicon technology inside the PC platform continues to evolve and deliver
twice the processor capability every 18–24 months. The electronics around the
processor is also evolving so that these increased capabilities can be used by
the basic PC platform. The PC platform itself is diversifying into multiple
usage models and application areas, all driven by customer demand. This all adds
up to a better user experience, all made possible by Intel continuing to prove
Moore's Law to be true.
To see and hear a webcast replay of Gordon Moore's Intel Developer Forum keynote
presentation, visit the IDF web site.
For more information on the technologies and happenings at IDF, read the
Top Story "PC Evolution Accelerates at Fall Intel Developer Forum"
in this month's issue of Platform Solutions.
To learn more about the platform technologies Intel is driving to keep pace
with Moore's Law, visit the Platforms and Technology pages in Platform Solutions.
|
