This is nothing new IBM has been working
on this since the 60's. It has been explained
there are difficulties in putting these chips
into devices.
FISHKILL, N.Y. -- Running one step ahead of an avalanche, IBM
Microelectronics
is driving its silicon-germanium (SiGe) heterojunction transistor
technology into
production through an
unprecedented network of foundry relationships. By licensing
powerful players in the
RF- and communications-IC industries, the computer giant has spread
its SiGe
BiCMOS process far beyond the walls of IBM--and may have sown the
undoing
ofgallium arsenide (GaAs) as an RF technology.
IBM's SiGe process adds 65-GHz SiGe heterojunction bipolar
transistors to a
relatively stock 0.4-micron CMOS formula. The combination permits
integration of RF
front ends with
intermediate-frequency, control-logic and, eventually, baseband
circuitry to produce single-chip radios, network interfaces and
the like. In communications, it permits the integration of
extremely high-speed signal paths with switching logic. And in
computing, the applications remain mostly unexplored but are
open-ended.
IBM developed the SiGe BiCMOS process at its facility here, where it
has a capacity
of 600 wafers a month. The process is being transferred to a
production line in
Burlington, Vt., in a move that will raise capacity by about a
factor of five.
Simultaneously, IBM has licensed a number of other companies to use
the process, its
libraries and some megacells.
"There are about five of us [licensees]," said Chris Henningsen,
director of marketing at
Harris Semiconductor. "Each one has paid a pretty steep membership
fee. There are
Harris, Hughes
Networking, National Semiconductor, NorTel and Tektronix, plus one
company that
has not been named publicly."
Each has its own uses for the technology. Harris, according to
Henningsen, is targeting
a major reduction in the chip count for PCS telephone handsets and
similar
digital-radio devices.
National Semiconductor is also pursuing RF integration. Hughes
is working on networking chips and NorTel on high-speed switching
equipment.
IBM itself has produced a number of important cells in the RF area,
including
multiplex/demux, D/A conversion and A/D
conversion. The company plans to market at least some of the cells
in
standard-product ICs later this year.
"The standard products aren't coming quite as fast as
advertised, [so] IBM needs some good partners," said Fred Zieber,
president of
Pathfinder Research (San Jose, Calif.). "IBM can do the
manufacturing, but they're
slim in terms of marketing because they haven't made a full
transition to a merchant
company.
"And this is not computer-related. This is mixed-signal and analog,
and that's something
with which they haven't had internal expertise."
IBM and its partners won't be alone in the market for long.
Contenders in Japan,
North America and Europe are readying the technology for production,
either
unilaterally or through licensing arrangements with Canadian startup
SiGe
Microsystems (see sidebar). The flurry of activity is a strong
indication that SiGe has
moved well past the R&D labs and into the product-planning phase.
Speed boost. At its core, silicon germanium involves speeding
circuit frequencies by
adding small amounts of germanium to bipolar junction transistors.
Today, that's done
using ultra-high-speed vacuum chemical vapor deposition about
two-thirds of the way
through the manufacturing process.
IBM claims it can achieve a 65-GHz fmax, compared with 15 to 25 GHz
for the
fastest silicon-only transistors.
"In a nutshell, SiGe improves bipolar to the point where there's no
discrepancy in
performance [compared with GaAs], with the
cost structure of VLSI silicon," said Derek Houghton, president of
SiGe Microsystems
(Ottawa), which was spun off from the Canadian National Research
Council 18
months ago.
"Going forward, there will be more emphasis on system-on-a-chip,
where the RF is
brought onto CMOS. That's where we're heading."
The potential impact of the technology is illustrated in the
performance of IBM's cells
and in licensee Harris' plans.
"The test group here developed multiplexer/demux circuitry that runs
at 5 GHz--twice
as fast as any commercial GaAs
device, at half the power," claimed Paul Cunningham, IBM
Microelectronics
product-line manager. "On the mixed-signal side, we have a 10-bit
D/A and a 6-bit
A/D in the fab now.
[The A/D] looks like it will perform 8 Gsamples/s. The device
actually simulates out
much faster than that, but we are making
an allowance for package losses."
While IBM will focus on a few standard products and a lot of custom
work with
specific customers, Harris is taking aim
directly at the PCS phone market.
"Today, what we can clearly do is put together the antenna front-end
circuits--the
low-noise amplifier and power amp--and
the up- and down-converters on a single die," Henningsen said.
"The IF, demodulator and filters could potentially go there as well.
"In addition, in an average handset design there are about 200
passive components of
various types, mostly resistors and
non-critical capacitors, that are relatively straightforward to
integrate in this process. By sweeping those up, we can convert
about 200 pieces and $45 in cost to about 40 pieces and $14."
While the short-term aim of IBM and Harris is improved power and
integration over
current GaAs RF devices, the stakes will
grow much higher. Both consider the SiGe BiCMOS process a major step
toward the
direct-conversion receiver: That is, they are developing a receiver
that converts directly
from an RF analog signal to a baseband digital one, without IF
stages.
"For the consumer market, direct conversion is something we want to
do and
something we are researching now," Henningsen
said. "Our gurus call it the holy grail of RF design."
"I personally believe that SiGe can do direct onversion," said IBM's
Cunningham. "The
big question is power."
Henningsen agreed. "Zero-IF direct conversion certainly makes sense
up through 1.9
GHz. And I think we see a path to 2.4
GHz. But 3 to 5 GHz--we're not sure how to do that yet.
"But even when it's possible, there are perception problems. People
have tried it in the
past, and it hasn't worked, so you have to convince them again. And
frankly, in the
PCS market
it's not a make-or-break issue right now. We are talking about
eliminating $4 to $5 in cost by taking out the IF stages. That's
small stuff compared with integrating a lot of the passive
components and the RF/IF."
Another opportunity the SiGe process opens for Harris is in
dual-band,
800/1,900-MHz phones. "If you do a dual-band phone, you have a
difficult challenge
on power consumption if you have to do dual front-ends," Henningsen
said. "With
SiGe, we can cover both bands with a single chip."
Similar impacts will show up in communications. NorTel is reportedly
working on very
fast crosspoint switch arrays. "In
the wireline area, there is a lot of ground to be explored in the
area above OC-48," Cunningham said.
Process particulars. IBM has developed its process essentially as an
added module to
an existing BiCMOS pro-cess. The company uses a patented
low-temperature film
epitaxy step, operating below 800C, to form the heterojunction
transistors.
The amount of germanium in the junction area has been a
critical issue, according to Cunningham. "As the amount of Ge
increases, you have to
stay below the Matthews and Blakely limit," or the device becomes
unstable.
"But if you look at the entire process, compared with our standard
silicon BiCMOS,
there is very little added complexity," Cunningham said.
Henningsen said the process has come up smoothly at Harris as well.
"We already had
a BiCMOS process," he said, "and the
SiGe has come up with no problems at all."
The sudden surge of development in SiGe could be the writing on the
wall for GaAs.
SiGe designers say that they get all of
the individual transistor speed of GaAs, but with far less noise,
much higher uniformity of performance across a wafer, much greater
thermal
conductivity and a far better cost structure.
Those comparisons have led SiGe proponents to predict that the
technology will
eventually drive GaAs into an ultra-high-speed niche. GaAs is still
too difficult for many
companies to master because of the inherent instability of arsenic,
which causes
threshold voltages to vary across the wafers, lowering yields. GaAs
devices are also
generally produced on older, 4-inch-wafer production lines.
"Unlike silicon, a high percentage of the cost for GaAs is the wafer
itself," said Zieber
of Pathfinder Research. "One of the
potential promises with SiGe is that you can get the cost down. You
can imagine a lot
of wireless-communications things if the cost is low enough."
Jerry |
