PORTLAND, Ore.—The odd thing about magnetic spin is that it does not displace atoms, allowing walled domains on hard disks to switch between "1" and "0" without the fatigue mechanisms that eventually wear-out flash bit cells. Unfortunately, solid-state nonvolatile memories like flash, ferroelectric and even experimental resistive RAMs have limited lifetimes, according to IBM, which claims its racetrack memory combines the advantages of solid-state memories with the access mechanism of a hard disk drive.
"The key advantage of racetrack memories is that they do not move atoms—that is why flash memory, ferroelectric and even resistive memories wears out, because they are disrupting the state of matter," said Stuart Parkin, a Fellow at IBM Research-Almaden in San Jose, Calif.. "But in racetracks we move no atoms—all we are doing is rotating the spin which does not interact or cause any fatigue or endurance issues. We can read, write and erase racetrack memories indefinitely."
The racetrack works by fabricating a nanowire loop on an insulating substrate. Along the length of the racetrack is a write head for imparting magnetic spin, a read head for reading-out spin states and a pulse-generator that shifts the bits to the next value along the length of the racetrack. In effect, each racetrack acts like a single track on a hard disk—except instead of spinning the disk to read the next value, the walls between the magnetic domains are shifted around the closed loop in response to current pulses.
Using a carefully constructed apparatus, IBM has recently been able to carefully characterize the exact mechanism by which domain walls can be moved around the track, gathering the parametric data together that will eventually enable IBM to commercialize racetrack memories. In a surprising finding, despite the fact that no atoms move around the racetrack, the movement of magnetic domain walls around the track nevertheless possesses inertia and momentum—as if they had mass.
"We now have a better understanding of how the electrons transfer momentum to the domain walls," said Parkin.
IBM 's racetrack memory moves magnetic domain walls along a nanowire where they "race" around the nanowire "track."
If racetracks exhibited true mass-less motion, that would have meant nearly instantaneous movement of domain walls around the racetrack in response to current pulses. But instead the researchers found a significant lag (nanoseconds) at the beginning of the pulse and an equally significant overshoot (microns) when stopping. Fortunately, the effects canceled each other out.
"What we found is that the domain walls take time to start moving at the beginning of the pulse, but that time is exactly compensated by the distance it takes them to stop moving at the end of the pulse—they still move the same total distance," said Parkin. "To move the domain walls a certain distance along the race track, all we have to do to is supply a current pulse whose length is proportional to that distance."
Next the group wants to demonstrate a complete horizontal racetrack with integrated read, write and shift circuitry all on a chip that is 10-times more dense than today's flash memories. Once the horizontal technology is successful, the group will turn to sinking racetracks vertically on chips as thick as 10 microns. Such vertical racetracks could be 100 times denser yet, putting them on par in density with hard disks, but on a solid-state platform.
Oddly enough, the recession can be good for technologies like this. Back in the boom times this kind of stuff would have been rushed into production before it was ready. On the other hand, for the last few years technology has been allowed to fully develop. Some things can't be rushed.
This same group was on the cusp of bringing magnetic bubble technology to fruition in 81ish. A big issue was scaling the very complex and exotic Liquid phase epitaxy process on expensive garnet wafers. Let's hope for no such problems here.
thanks for nice research report. It will be even better to attach a link to either some recent publications or a report containing detailed work. I actually did not understand "The key advantage of racetrack memories is that they do not move atoms—that is why flash memory, ferroelectric and even resistive memories wears out, because they are disrupting the state of matter," part very much.
Racetrack memories work like a shift register--only with magnetic domains on a nanowire doing the shifting. The cool thing about the physics, is that no matter which way current pulses the nanowire--up or down--the imparted momentum pushes all the existing domains on the nanowire along in the same direction. The domains appear to be "pushed" along the wire, but of course is just the spins of the atoms that are moving--like the "wave" at the ballpark, where everyone stays in their seat, but just stands up at the right moment, thereby presenting the illusion of a moving wave. Likewise, magnetic domains are shifted around the nanowire loop even though the atoms stay fixed in place.
Racetrack memories have been an intense research area for IBM Fellow Parkin since before 2007 when I first discovered his work. Since then, he has perfected most of the necessary components--read-head/write-head/shifter--and is now entering the "process integration" step, in which IBM will attempt to fabricate all the separate components on a single CMOS chip.
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for today’s commercial processor giants such as Intel, ARM and Imagination Technologies.