|
Researchers from Rice University and the University of California,
Los Angles unveiled a new data-encoding scheme that slashes more than 30% of
the energy needed to write data onto new memory cards that use phase-change memory
(PCM)—a competitor to flash memory that has big backing from industry
heavyweights.
The breakthrough was presented at the IEEE/ACM Design
Automation Conference (DAC) in San Francisco by
researchers from Rice
University’s Adaptive
Computing and Embedded Systems (ACES) Laboratory.
PCM uses the same type of materials as those used in
rewritable CDs and DVDs, and it does the same job as flash memory—the mainstay
technology in USB thumb drives and memory cards for cameras and other devices. IBM
and Samsung have each demonstrated PCM breakthroughs in recent months, and PCM
is ultimately expected to be faster, cheaper, and more energy-efficient than
flash.
“We developed an optimization framework that exploits
asymmetries in PCM read/write to minimize the number of bit transitions, which
in turns yields energy and endurance efficiency,” said researcher Azalia
Mirhoseini, a Rice graduate student in electrical and computer engineering, who
presented the research results at DAC.
In PCM technology, heat-sensitive materials are used to
store data as ones and zeros by changing the material resistance. The
electronic properties of the material change from low resistance to high
resistance when heat is applied to alter the arrangement of atoms from a
conducting, crystalline structure to a nonconducting, glassy structure. Writing
data on PCM takes a fraction of the time required to write on flash memory, and
the process is reversible but asymmetric; creating one state requires a short
burst of intense heat, and reversing that state requires more time and less
heat.
The new encoding method is the first to take advantage of
these asymmetric physical properties. One key to the encoding scheme is reading
the existing data before new data is written. Using a combination of
programming approaches, the researchers created an encoder that can scan the “words”—short sections of bits on the card—and overwrite only the parts of the
words that need to be overwritten.
“One part of the method is based on dynamic programming,
which starts from small codes that we show to be optimal, and then builds upon
these small codes to rapidly search for improved, longer codes that minimize
the bit transitions,” said lead researcher Farinaz Koushanfar, director of
Rice’s ACES Laboratory and assistant professor of electrical and computer
engineering and of computer science at Rice.
The second part of the new method is based on integer-linear
programming (ILP), a technique that can find optimal solutions. The more
complex the solution, the longer ILP takes to find the optimal solution, so the
team found a shortcut by using dynamic programming to create a cheat sheet of
small codes that could be quickly combined for more complex solutions.
Research collaborator Miodrag Potkonjak, professor of
computer science at UCLA, said the team’s solution to PCM optimization is
pragmatic.
“The overhead for ILP is practical because the codes are
found only once, during the design phase,” Potkonjak said. “The codes are
stored for later use during PCM operation.”
The researchers also found the new encoding scheme cut more
than 40% of “memory wear,” the exhaustion of memory due to rewrites. Each
memory cell can handle a limited number of rewrite cycles before it becomes
unusable.
The researchers said the applicability, low overhead and
efficiency of the proposed optimization methods were demonstrated with
extensive evaluations on benchmark data sets. In addition to PCM, they said,
the encoding method is also applicable for other types of bit-accessible
memories, including STT-RAM, or spin-transfer torque random-access memory.
Source: Rice University