IBM Pares Speed Gap In Memory Circuitry
(Design for Data Storage May Be in Use Next Year; Performance Could Double)
By DON CLARK
February 14, 2007
International Business Machines Corp. is
claiming a breakthrough in developing circuitry to store data on future
The big computer maker said its approach -- based on exploiting the most
widely used memory technology in a new way -- could triple the data stored
on a typical microprocessor with a resulting doubling of computing
"We think this is the next big thing in getting more system performance,"
said Lisa Su, IBM's vice president of semiconductor research and
development.IBM plans to use new memory technology, shown in this diagram of
a prototype chip.
Microprocessors, the calculating engines for computers, increasingly come
with storage circuitry to minimize the delays associated with fetching data
from external memory chips. This "cache memory" typically uses a kind of
circuitry used on chips called SRAMs, or static random-access memories.
SRAMs are fast but require six transistors to store a single bit of data.
The more widely used chip known as dynamic random-access memories, or DRAMs,
only need one transistor and another component, a capacitor, to store a bit.
But DRAMs, though they can store more data in a smaller space, have
generally been considered too slow for cache memory.
IBM researchers are discussing their progress in closing the speed gap at a
conference in San Francisco today.
Exploiting a manufacturing technology called silicon-on-insulator, the
company has developed unusually fast DRAM circuitry for use as cache memory.
Subramanian Iyer, a director of IBM's manufacturing-process development,
estimates it takes 1.5 nanoseconds -- or billionths of a second -- to fetch
data from its enhanced DRAM technology, compared with 10 to 12 nanoseconds
for conventional DRAMs and 0.8 to 1 nanoseconds for SRAMs. Mr. Iyer said
three times more data can be stored in the same amount of space by switching
from SRAM to DRAM circuitry; he expects the technology to be incorporated on
microprocessors that will be manufactured next year using a new production
Such benefits could help IBM's Power microprocessors in a performance race
with chips from Intel Corp. and others. But Shekhar Borkar, the director of
Intel's microprocessor technology lab, said extra manufacturing costs
associated with using DRAM circuitry could outweigh the benefits.
IBM is a technology partner with Advanced Micro Devices Inc., an Intel rival
that could benefit from the computer maker's memory research. Meanwhile,
other alternatives to SRAM for cache memory are also being studied.
Innovative Silicon Inc., a start-up, has been promoting a technology it
calls Z-RAM that stores data using a single transistor. Its licensees
An AMD spokesman said it is "evaluating a number of new and emerging
technologies" for cache memory.
IBM drives road to denser CPU memory
Rick Merritt / EE Times
(02/14/2007 12:00 AM EST)
SAN FRANCISCO, Calif. — IBM Corp.
will detail a method for tripling the amount of memory on a
microprocessor, potentially doubling its performance. By
combining techniques in process and circuit design, IBM believes
it can put as much as 48 Mbytes of fast DRAM on a reasonably
sized CPU when its 45nm technology becomes available in 2008.
IBM's upcoming Power6 CPUs use 8 Mbytes SRAM cache. Rival Intel
Corp.'s Itanium processors use as much as 18 Mbytes.
"Processors are definitely cache starved, and as you go more
towards multi-core processors, the need for memory integration
becomes more acute," said Subramanian Iyer, a distinguished
engineer and director of 45nm technology development at IBM.
"There are some server chips that could not be made without this
technology," he added.
In a paper at the International Solid State Circuits Conference
here Wednesday (Feb. 14) IBM will describe a 65nm prototype
embedded DRAM with a latency of just 1.5 ns and a cycle time of
2 ns. That's an order of magnitude faster than today's DRAMs and
competitive with SRAM that is typically used for microprocessor
"To put 24-36 Mbytes of memory on a chip, you would need a
600mm-squared die today. Using this technology you could put
that much memory on a 300-350mm-squared die," Iyer said.
IBM expects to use the technique on its future Power and Cell
processors as well as have it available for its ASIC customers.
"It's being defined in a way that it can be part of our standard
45nm process technology," Iyer said.
IBM combined two advances to enable the new memory integration.
The company found a way to migrate its deep trench technology
used for DRAMs from CMOS to its silicon-on-insulator (SOI) logic
process. In a paper last December, IBM described that work that
involved suppressing the floating-body effect in SOI.
"Our entire processor road map is based on SOI," said Iyer.
New circuit designs use short bit lines to eliminate the need
for sense amps that detect voltage differences between the bit
lines and a capacitor, a process that makes DRAMs relatively
slow. The new design uses a three-transistor micro-sense amp
that lets voltage current directly drive transistor gates.
IBM used embedded DRAM in a custom processor designed for its
high-end Blue Gene/L supercomputers, but has not been able to
use the technology in mass market computer chips to date. "This
is 100 percent mainstream and we expect to get it in products in
2008," Iyer said.
Intel and other chip makers are investigating using the floating
body cells to store charge as one alternative for embedded
memory. Other chip makers are researching stacking memory and
processor dice in multi-chip modules.
Intel archrival Advanced Micro Devices co-develops process
technology with IBM and could use the embedded DRAM technology
as a way to compete with Intel.
Floating-body memories gain ground
David Lammers / EE Times
(07/17/2006 9:00 AM EST)
Austin, Texas -- Semiconductor researchers
are pursuing an emerging memory type that uses the floating-body effect
of silicon-on-insulator (SOI) technology as the memory storage
mechanism. At last month's Symposium on VLSI Circuits, Renesas
Technology Corp. and Toshiba Corp. described progress with their
floating-body memory (FB-RAM) development efforts, aimed at embedding
fast and dense DRAM on logic chips that require lots of on-board memory.
Earlier this year, Advanced Micro Devices Inc. agreed to license the
Z-RAM (zero capacitor, one transistor) technology from Innovative
Silicon Inc., a Santa Clara, Calif.-based startup with its engineering
team in Lausanne, Switzerland.
Floating-body memory is still years away from commercial use, however.
"Companies may do SOI-based products, but I don't think they will put
floating-body RAM on SOI until at least the 25-nm node," said Sreedhar
Natarajan, president of Emerging Memory Technologies Inc. (Ottowa) and
the author of a book on SOI chip design.
The FB-RAM is "basically a small, cheap embedded DRAM--it is not an SRAM
replacement," said Suresh Venkatesan, a director of silicon technology
solutions at Freescale Semiconductor Inc. "With any embedded DRAM, you
have to go through the process of seeing where it fits. The question is,
now that it is smaller and cheaper, does it open up more application
spaces than the old embedded DRAM?"
Venkatesan foresees a decade's worth of development ahead. "Any new
material or technology, in my opinion, will take eight or 10 years [to
perfect]," he said.
FB-RAM does away with the capacitor used in conventional DRAM bit cells
built in bulk silicon. In bulk CMOS, the charge that forms a
transistor's body is tied to a fixed voltage. In SOI, the untied body is
"floating" in silicon above the thick oxide layer. To make the floating
body behave like a capacitor, a carefully controlled voltage is applied
on both sides of the body.
For companies that start out with SOI wafers for their high-performance
processors, such as AMD, IBM or Freescale, the FB-RAM approach has
several advantages: fast read and write times, and a cell size smaller
than embedded DRAM and about one-fifth that of a six-transistor SRAM.
With the trench and stacked capacitors of conventional embedded DRAM
becoming more difficult to build, proponents claim that FB-RAMs are
cheaper, and require few additional mask layers.
There are challenges. Since the floating body is not tied to a fixed
voltage, the stored charge can easily fluctuate if the reference voltage
is not strictly controlled. As CMOS scales to 45 nanometers and below,
controlling transistor and voltage variation is becoming more difficult,
including variations in FB-RAMs. If the voltage fluctuates, plus or
minus, by 0.5 V, the FB-RAM transistor can lose charge.
Like SRAMs and DRAMs, FB-RAMs are volatile--data is lost when the system
powers down, so FB-RAMs must be refreshed every few hundred
milliseconds. But they do not need to be restored in the same way as
DRAMs, which require data to be restored after the destructive DRAM read
At the Symposium on VLSI Circuits, Toshiba engineer Takashi Ohsawa said
the company's 128-Mbit floating-body cell (FBC) test chip demonstrated a
nearly 100 percent bit yield. The cell size of 0.17 square microns is
half that of comparable embedded-DRAM bit cells. "Scaling is easier than
with conventional embedded DRAM," Ohsawa said. "With DRAMs, the process
cost is getting high because of the deep-trench capacitor." With FBC,
however, "it is important to have an accurate reference voltage." In the
Toshiba FBC, the write time was consistently measured at 10 nanoseconds
Also at the symposium, Renesas discussed an enhanced twin-transistor RAM
based on an SOI technology. Though the design has a larger cell size
than the single-transistor FB-RAM approach, Renesas said that it is less
costly to build, with fewer process changes. Also, the company claims
that the so-called eT2RAM is better-suited to modern system-on-chip (SoC)
solutions, which rely on multiple voltages and other power-management
schemes to keep power consumption under control. "This could be a
mainstream memory for advanced SoCs," said Kazutami Arimoto, a Renesas
The company recently fabricated a 4-Mbit test chip using a 90-nm SOI
process, and a full evaluation is under way. With the two-transistor
architecture, the eT2RAM operated successfully with a 0.5-V power
supply, achieving a respectable 56-MHz operation at such a low operating
If a high-k dielectric and metal gate electrode were used in the
transistor, the cell size could be scaled further--partly because the
high-k dielectric can be physically thicker and transistor variability
can be better controlled, Arimoto said.
Since Renesas--ranked as the world's largest microcontroller
vendor--does not have a line of high-performance MPUs that use SOI
technology, one challenge is to find a home for its SOI-based eT2RAM.
In May, Emerging Memory Technologies Inc., which provides memory design
services and intellectual property, took a license. EMT is betting that
SOI will expand beyond the high-performance MPU niche into the broader
semiconductor market, company president Natarajan said.
"Customers will eat the area penalty with the Renesas twin-transistor
approach if the thing actually works," he said. "The data retention is
better, the performance is better, the yield is better and it doesn't
need a high-voltage [prevention] mask. The high voltages also become a
Renesas is working to adapt the twin-transistor RAM process to an SOI
process at a major foundry in Taiwan, so that potential customers can
Meanwhile, Venkatesan said Freescale has been evaluating the Innovative
Silicon approach to FB-RAM, adding that the company has not taken a
license from ISI at this point. "We continue to evaluate the technology
for our own set of applications. ISI has some ideas about how certain
things can be done to avoid some of the issues that one could
potentially face with a memory type of that nature," he said.
The drawback of the FB-RAM is that it is fundamentally available only on
an SOI substrate, which limits the number of applications. Network
processors, for example, are not well-suited to embedded DRAM,
Venkatesan said, because many NPUs are crammed onto line cards, making
it difficult to partition the memory. For high-performance CPUs, which
face memory limitations going forward, the FB-RAM could not be used to
replace SRAM in the Level 1 cache, which must run at multiple gigahertz.
"With floating-body RAM you have to look for a system architecture that
makes it attractive to have large amounts of memory on-chip that doesn't
need to be accessed within the clock cycle, or which can be accessed
with much higher latency," Venkatesan said. "One example might be
pulling in some off-chip DRAM back onto the processor. One has to take a
look at what this technology brings to the table and then do things
differently in terms of how you partition the memory. That's a process
of understanding that takes time."
FB-RAMs are in the early stage of development, and it will take time to
explore the reliability issues, he added. "The memory incumbents [DRAM,
SRAM and flash] have been around for three decades or more, and we
understand their reliability. With any new technology, there is the
whole piece of understanding the reliability issues and scaling path,"
he said. "It took Freescale 10 to 12 years to get its MRAM to the point
of commercial introduction [see July 10, page 1]. It took us 10 years to
bring SOI to the market. It took us 10 years for copper interconnects."
Venkatesan thinks FB-RAM faces the same development curve.
Table, Schematic, Layout, and Cross-section of TTRAM and FB-RAM