IBM has published a paper on drift-tolerant multilevel phase-change memory. That may not sound exciting, but it could mean memory chips that have all the benefits of flash memory, but work far quicker and last far longer. If that proves the case, the chips could become useful for business machines.
The basic concept of phase change memory is that data is stored via a type of material named chalcogenide compounds. To put things in a very simplified form, the application of heat switches the material from a crystal state to an amorphous state: the technique is also used in rewriteable optical disks.
What makes chalcogenide compounds so suited to the task is that the difference between the two states is very distinct, to the extent that it’s possible to identify two “stopping off points” during the transition. This means a total of four identifiable “positions” for each storage unit, thus increasing the amount of data that can be stored.
Another benefit is that the process doesn’t require old data to be erased before new data is written to the same space, thus speeding up performance.
The problem to date has been that not only does the level of resistance of the material drift randomly over time, but the amount of drift varies depending on the particular state (crystal or amorphous) that a storage point is in. The difficulty of keeping track of this drift and adjusting to meet it has meant the material has a limited lifespan when it comes to being used for data storage.
The image above, taken from the IBM paper, shows the effects of the drift on a collection of 200 cells: figure a is the point at which data is written, figure b is 40 microseconds later (forty-millionths of second), figure c is one thousand seconds later and figure d is 46,000 seconds (just under 13 hours) later. Again put very simply, the practical effect of what’s represented by the collapse of the red line is increasing difficulty in reliably reading the stored data.
The IBM paper discusses the discovery that although the beginning of such drift is random, the way in which it spreads across the material can be predicted with an algorithm. That makes the adjustment process simpler, meaning the material can be used longer before keeping track of the drift becomes impractical.
The upshot is that phase change memory using this system should be able to last long enough to be written to five million times. By comparison, existing flash memory usually lasts for between 5,000 and 10,000 write cycles when produced in a way that makes it affordable for consumer products.