Navigating The DRAM Market, Technology Maze -- Reaching a successful end requires walking a fine line between investing too much too soon or too little too late.
Date: 08/18 23:44 EST
Aug. 18, 2000 (Electronic Buyers News - CMP via COMTEX) -- The turn of the
century is shaping up as a time of plenty for DRAM manufacturers and their
customers. Desktop and notebook PC markets continue to grow at a healthy rate,
and the server market is exploding thanks to the Internet. Faster processors and
bandwidth-hungry applications are stoking demand for memory.
However, the emergence of different types of DRAM for various applications
complicates the quest for market and mind share. Manufacturers are forced to
formulate capacity, product-development, and production plans based on forecasts
of which technology will take off.
But the market landscape is changing, perhaps as rapidly.
"It's not like the old days now, since four or five suppliers are making 90% of
all DRAMs," said Jan du Preez, vice president of memory products at Infineon
Technologies Inc., San Jose. "You can't hire and fire suppliers anymore.
Strategic alignments and relationships are forming; road maps are being
aligned."
In the past couple of years, Micron Technology acquired Texas Instruments' DRAM
business and Hyundai acquired LG Semiconductor. Next year, NEC and Hitachi are
merging their DRAM operations.
Hitachi, Hyundai, Infineon, Micron, NEC, and Samsung control the lion's share of
the market. Other vendors are cultivating niches and may not be able to
establish or maintain positions in both double-data-rate (DDR) SDRAM and Rambus
(RDRAM) technologies.
Although DRAM makers are enjoying prosperity, the story was different not so
many quarters ago. Facing perennial market volatility, DRAM makers were hesitant
to invest upward of $1 billion for each new wafer fabrication facility.
Infineon is building a 300-mm factory in Dresden, Germany, du Preez said.
"Building the factory will take a year. We'll start our initial ramp in Q4 2001,
and the real capacity will come on line around the end of 2002," he said.
DRAM demand is soaking up capacity, said Steve Cullen, an analyst at In-Stat
Group, Scottsdale, Ariz. "For the next couple of years, demand will exceed
supply, and that's welcome news for the DRAM guys that've been hurting since
1995.
"We think the market just entered a real shortage-approximately 3% to 5%
unfilled demand in the most recent quarter," said Reiko Soga, senior product
marketing manager at NEC Electronics Inc.'s North American memory business, in
Santa Clara, Calif. Unfilled demand could increase to 10% in the current
quarter, she said.
Shortages are occurring in certain configurations, "and some OEMs, when they
can't find parts, are willing to pay anything," Infineon's du Preez said. "But
vendors can't seriously abuse the market."
NEC expects the rates of DRAM consumption by application to change significantly
in the next few years. "Historically, PC applications dominated DRAM
consumption, but communications and consumer portions may exceed the PC segment
within the next few years," Soga said.
DRAM suppliers must have strategies and product-development plans to support
emerging technologies, she said.
Makers must walk a fine line between investing too much too soon or too little
too late.
"The DRAM business is always tough. It's never in balance," said Cecil Conkle,
assistant vice president of DRAM marketing at Mitsubishi Electronics America
Inc., Sunnyvale, Calif. "There are always issues to deal with, whether it's a
shortage of supply, demand is low, or pricing is too high for customers or too
low for us. DRAM manufacturers are very competitive, and there's always a war of
one kind or another."
Sorting out DDR, Rambus
The DRAM industry is preoccupied with the DDR-Rambus battle.
"We have to make investment decisions with hundreds of millions of dollars,"
said Bob Fusco, DRAM marketing manager at Hitachi Semiconductor (America) Inc.,
San Jose. "We have to be sure that the markets we select are going to take off.
For example, both Rambus and DDR are going to go along together for a while, but
the market will ultimately decide."
There are lots of unknowns regarding Rambus and DDR, said Jaime Stitt, business
development manager of DRAM memory products at Toshiba America Electronic
Components Inc. (TAEC), Irvine, Calif. "There are benefits to both, but they
bring complexities in packaging and testing that we haven't seen before. There's
probably a home for both, but the industry needs an infrastructure to guarantee
success."
Today, the dominant memory device is the 100-MHz PC100, but that's
changing-fast. "PC133 started picking up heavily in 2Q based on a lot of
speculative buying," said Victor De Dios, president of De Dios & Associates,
Newark, Calif.
"The true conversions are beginning now and will occur throughout the second
half of this year," De Dios said. "It should be a fast transition-from 5% of the
market in 1Q to 33% by 4Q. By 1Q next year, the PC133 market will be larger than
the PC100 market."
PC100 and PC133 SDRAMs currently account for 85% of the market, EDO about 8%,
and Rambus and DDR together around 6% to 8%, according to Farhad Tabrizi, vice
president of marketing at Hyundai Electronics America, San Jose.
"We're seeing the majority of graphics applications moving to DDR," he said. "In
low-end PCs, we expect the move to DDR to start in the fourth quarter. In
high-end, Intel-based PCs, we'll see some Rambus."
"In the server market, which is 20% of the total DRAM market, it's all SDRAM,
moving to DDR in the first half of next year," Tabrizi said. "Workstations use a
combination of EDO and SDRAM."
Some server makers are interested in DDR, others in Rambus, depending on how
their systems are put together, TAEC's Stitt said.
"The logical progression from PC133 and fast SDRAMs is to DDR DRAMs," according
to a recent report from IC Insights Inc., Scottsdale, Ariz. "DDR further
increases the performance offered by PC133.
"It increases bandwidth by clocking data on both edges of the DRAM clock. ...
DDR is quite compliant with existing PC133 SDRAMs. Migrating from PC133
architecture, DDR requires minimal changes in system bus width and module pin
count. Compared with Rambus DRAMs, DDR production costs are inexpensive,"
according to IC Insights.
IC Insights predicts that PC133 will represent 37% of the global DRAM revenue
this year, 55% next year, and 44% in 2002. The research firm pegs DDR at 6% this
year, 13% in 2001, and 23% in 2002.
While some see PC133 as a pit stop on the way to DDR, others are more
sanguine.
NEC is focusing on 128-Mbit PC133 production for the rest of this year, Soga
said. Meanwhile, the NEC-Hitachi joint venture is developing a 0.13-micron
process that it hopes to sample in next year's second quarter. NEC plans to add
12-in. wafer capacity in 2002 or 2003.
"We believe that demand for SDRAMs will remain strong throughout 2001, " Soga
said. "Under current short-supply market conditions, we have the opportunity to
establish and improve long-term strategic partnerships with key customers by
providing strong product support," she said. "Also, it's critical that we
achieve the optimum product mix, execute scheduled die shrinks, and also
maintain maximum volume production."
In May, Micron Technology Inc., Boise, Idaho, began sampling 133-MHz 2-2-2
SDRAM-registered and unbuffered DIMM modules in densities ranging from 32 to 512
Mbytes.
The 2-2-2 is a faster version of the prevailing PC133 3-3-3. The numbers refer
to timing specifications in clock cycles. Column-access (CAS) latency is the
time from read command to data out; row-address strobe (RAS) to CAS delay is the
time from the active command to the read command, and RAS precharge is the time
from precharge command to active command.
Micron was the first DRAM manufacturer to meet the Intel 133-MHz SDRAM device
specification goal of 2-2-2.
Enhanced Memory Systems Inc., Colorado Springs, Colo., this month is launching a
150-MHz CAS 2 (2-2-2) device.
Latency concerns
The main advantage of 2-2-2 SDRAMs compared with 3-3-3 devices is improved
lead-off latency, said Chris Johnson, SDRAM applications engineer at Micron.
"Both speed options provide the same peak burst rate, but PC133 2-2-2 offers a
lead-off latency of only 30 ns, which is a 34% improvement over the 45-ns
lead-off latency that is available from PC133 3-3-3 SDRAM," he said.
The faster SDRAMs provide a quick and easy way to boost performance without a
design change, according to Kit Fawcett, DRAM product marketing manager at
Micron.
Executives in the DRAM industry agree that latency-the lower the better-is a
critical measure of performance. There's disagreement, however, over which
technology offers the lowest latency.
"Rambus uses a 16-bit bus, where-as DDR SDRAMs work with a 64-bit bus, " said
Dave Bondurant, vice president of marketing at Enhanced Memory Systems. "That's
the same bandwidth, at a lower clock rate for DDR [200 MHz vs. 800 MHz]."
"Cramming so much data through a narrow Rambus bus adds latency," Bondurant
said. "And the latency is more significant than the fact that Rambus moves words
four times faster."
A DRAM's latency is determined by the speed of the memory core, countered a
Samsung spokesman. Since commodity DRAMs all use the same memory-core
technology, any differences in latency are due to the different speeds of their
interfaces.
An 800-MHz RDRAM has a latency of 38.75 ns, compared with 40 ns for the PC100
and 45 ns for the PC133, the Samsung spokesman said. System latency is
determined by adding external address and data delays to the component latency,
and total SDRAM system latency is 90 ns for PC100 and 75 ns for 133-MHz SDRAM.
RDRAM system latency is 70 ns.
The latency problem plagues DDR as well as Rambus, according to Enhanced Memory
Systems' Bondurant. Doubling a data rate from 133 to 266 MHz doesn't have much
impact without a concomitant improvement in latency. The company said it will
lower latency with a caching scheme that cuts page-access times in half.
Enhanced Memory is hoping that the scheme will be adopted as part of the
standard for DDR2, the next generation, which SDRAM advocates say will match
Rambus' clock speed.
But there's more to life in the DRAM world than latency.
Intel Corp., Santa Clara, Calif., conducted benchmark tests showing PC133 SDRAMs
and Intel's 815E chipset wiping up on or at least edging RDRAMs and the Intel
820 in nine of 11 head-to-head match-ups, according to recent industry reports.
"The benchmarks showed a 2% performance advantage for PC133 over Rambus," said
Bert McComas, an analyst at InQuest Inc., Gilbert, Ariz. "Posting the benchmarks
on their Web site amounts to Intel officially admitting that the PC133 is faster
[than Rambus, which Intel has championed].
"There's still a lot of propagation of old thoughts and ideas that Rambus is
better, but the bulk of the market's over that," McComas said. "Not one major
OEM has made the decision to go with Rambus."
DDR isn't ready yet, according to McComas. "Best case, it'll ship in 4Q to OEMs
that are willing to take risks," he said.
Ready or not, the high end of the DRAM market has its appeal.
In July, Toshiba launched a line of 256-Mbit DDR SDRAMs, DDR fast-cycle RAMs
(FCRAMs), and Rambus RDRAMs. The plethora of parts is intended for workstations,
servers, and high-end PCs as well as image-processing and network applications.
They are manufactured on a 0.175-micron process that enables Toshiba to support
a variety of applications, Stitt said.
The DDR FCRAMs were originally developed by Fujitsu Microelectronics Inc., San
Jose. They lower power consumption by narrowing the memory-active areas and
incorporating a proprietary core technology that can achieve a 30-ns
random-access cycle time.
FCRAMs are suitable for applications that need high DRAM den-sities and
near-SRAM cycle performance, Stitt said.
A spokesman at Fujitsu described the company's FCRAMs as "sector-specific,
application-specific, or semi-custom memory devices that move data much faster
than the jellybean commodity versions."
Early this year, Fujitsu launched a 64-Mbit memory with an SDRAM interface for
graphics applications. It followed that part with a 64-Mbit memory for set-top
boxes, handheld devices, and other consumer applications, and in July, Fujitsu
introduced a 16-Mbit SDRAM with an SRAM interface designed for use in cell
phones. The Fujitsu spokesman said Toshiba will manufacture second-generation,
128-Mbit FCRAMs.
"Fujitsu wants to move away from the extremely cyclical commodity-memory
business, but also realizes that multimedia, communications, and networking
products need different kinds of memory capabilities, at different speeds and
bandwidths that include high-speed interfaces and specialized packages," the
spokesman said. "Applications are different and diverse enough that customers
are willing to pay the premium [as much as $25 to $30 each] for the specialized
memory-at least for now."
Hedging bets
As the DRAM industry leader, Samsung has the resources to bet on all the
metaphorical horses. In February, it showed an 800-MHz, 128-Mbit RDRAM in an
Intel 820-based system and a system with 1 Gbyte of SDRAM consisting of 36
256-Mbit PC133 components. The 256-Mbit chips are manufactured on an 0.18-micron
process and have the same package footprint as 64- and 128-Mbit SDRAMs.
In April, Samsung launched a 512-Mbit DRAM based on 0.12-micron process
technology. The part uses the same packaging as 256-Mbit DRAMs and can support
both SDRAMs and DDR. Thirty-six 512-Mbit components can be configured into a
2-Gbyte module.
Rambus Inc., Mountain View, Calif., in July announced support for a 1, 066-MHz
RDRAM, which boosts performance by a third, compared with its 800-MHz, 1-Gbit/s
chips. The technology is a simple bin split of the 800-MHz RDRAM and is aimed at
applications such as OC-192 line cards, video consoles, HDTVs, set-top boxes,
digital-video recorders, and information appliances, as well as graphics and
communications applications. It provides 2.1 Gbits/s of bandwidth from a single
device on a Rambus Channel.
Hyundai, Infineon, NEC, Samsung, and Toshiba should all unveil devices based on
the technology by next year, according to executives at Rambus.
Easy for them to say. "DRAMs are extremely complex products that take a long
time to design," Infineon's du Preez said. "You need bright engineering talent,
and that requires an investment. Manufacturing is done in factories that cost $1
billion or more, and building a factory takes two years."
DRAM manufacturers must be aggressive in risk management and confident in what
they are doing, du Preez said. "You need a secure customer base and $3 billion
for a start-up investment, and the market's very unpredictable. That's why there
have been no new entrants in recent years.
"The firms that are in the market have invested billions, are successful, and
are going to remain in the business," du Preez said. "Some years you can make
significant profits, and this is one of them."
ebnonline.com |
