|
Koelmans and Sebastian in the lab where the
new memory design was developed |
IBM researcher Wabe Koelmans had the first significant eureka
moment of his young, promising career on Saturday, 14 December 2013.
The Dutch
scientist is focused on emerging memory technologies based on what are known as
phase change materials. Such materials can be used to store data quickly,
simply by applying electrical pulses to the material, which is sandwiched
between two metal electrodes.
Koelmans
started his experiment the prior evening at IBM’s Zurich Lab
and then went home for the night to monitor the results from the comfort of his
apartment. By Saturday he and his colleague Abu Sebastian, were already seeing significant
progress.
After
Koelmans made a small tweak to the experiment they exchanged some messages, and
similar to waiting for a movie to finish downloading, they anxiously stared at
their screens and waited.
“It was extremely addictive to check
the experiment’s progress, so we were both very
excited and logging in regularly. I ran over 60 experiments, each taking half an
hour, in that weekend, but it was worth it,” said Koelmans.
By Sunday
morning, with a cup of coffee in hand, Koelmans saw the remarkable — drift was
virtually non-existent. The significance of this requires a bit of explaining.
An
Introduction to Drift
Phase change
memory (PCM) devices have been investigated since the early 1970s and in the
past 12 months IBM scientists have made tremendous progress with this
technology publishing a number of milestone papers which demonstrate multiple
bits per cell, making the technology extremely competitive with Flash.
But PCM doesn’t come without some drawbacks — a primary culprit being resistance drift. Drift is the change in resistance of the stored
levels over time. Essentially the data moves causing your text document or your
photo to eventually become corrupted and unusable — a very bad characteristic
for a storage technology.
|
Fig B: The desired I–V characteristics corresponding to the
phase-change and projection
segments are shown schematically Credit: Nature Communications |
A significant breakthrough to eliminate drift came a
couple of years back when IBM scientists from Zurich and Yorktown Heights, NY proposed
a novel device concept where the physical mechanism of writing is decoupled from the
read process. To use an analogy, during the write cycle the electric pulse goes
straight down through the material like water washing over a boulder in the
middle of a river. But as in figure B during the read, the pulse goes around the material, or
the boulder, decoupling the write and the read. This allows storing information
without suffering from the drift that normally comes with it.
“Over the years,
IBM scientists have developed powerful techniques mostly based on read/write
methodology and signal processing to counter resistance drift in order to
facilitate multi-bit storage," said Abu Sebastian, one of the IBM co-inventors of the concept comments. "However, a device-level solution with substantially
lower drift will further improve the error-rate performance and could enable
the storage of a higher number of bits per device using the same techniques.”
Despite the
promise of the concept, a physical realization of such a device with a
conclusive experimental demonstration of all its benefits was still missing.
Eureka 2.0
During a period of eight months, Koelmans and his
colleagues continued optimizing the design and materials to achieve fast write
speeds and high endurance, attributes that make PCM a compelling memory
candidate for high-throughput applications in the cloud.
Then came Monday, March 30, 2015. On this fateful spring day,
Koelmans had his second significant eureka moment: near elimination of noise.
In PCM devices, besides drift, the resistance levels
also experience random fluctuations. This low-frequency noise (1/f) behavior is
a nightmare for multi-level storage. Drift having been dealt with, this noise is
arguably the second largest challenge that limits the number of bits one could
store in a single PCM device. Koelmans observed almost complete elimination of
this 1/f noise in these devices.
“It’s easy to focus on the near-elimination of resistance drift, but this is
only one of the benefits. The two others: much reduced 1/f noise and much
reduced - and predictable - temperature dependence are also very important. In
the end, you need a stable signal (resistance) over time and temperature and a
good signal-to-noise ratio to facilitate multi-level PCM,” said Koelmans.
|
Schematic 3D view of projected
phase-change memory devices with lateral geometry.
Credit: Nature Communications
|
The next
morning Koelmans ran into the office of team leader and IBM Fellow Evangelos Eleftheriou,
who was astonished, and agreed that they needed publish right away.
Five
months later Koelmans and the team delivered. Today in the peer-reviewed journal
of Nature Communications (10.1038/NCOMMS9181) this new, radical memory cell design
is being reported on for the first time which the team calls projected phase
change memory, which features virtually no drift and very little noise. The
paper introduces the concept, along with the design, fabrication, and
simulation of such devices with the results.
IBM scientists believe that projected phase change
memory devices could also play a key role in future non-Von Neumann computing
paradigms such as brain-inspired, neuromorphic computing.
Labels: memory, nature communications, PCM, phase change memory, zurich