Phase-Changing the Game

How much memory is enough for a computing system? This is something of a trick question, because historically, the answer has always been “just a little bit more.” Not so very long ago, people would have said that 64 KB, 640 KB, or 1 MB (gasp!) would be more than any application would ever need. But as memory technology has advanced and declined in cost, people have invariably found a way to use all the memory that they could get their hands on, and then pine for more.

In recent years this problem has become much more acute with the rise of interest in data-intensive applications, especially artificial intelligence (AI). AI algorithms involve massive numbers of computations that need to be performed rapidly. This means that a tremendous amount of data needs to be moved between the memory and processing units, which is very slow and requires a lot of energy. These factors make the prospect of scaling up present bleeding-edge AI applications very challenging and unsustainable.

As AI continues to evolve and permeate ever more aspects of our lives, these problems will only grow larger. A paradigm shift may be needed in computing, away from the traditional architectures that have served us so well in recent decades, towards a hardware platform that is designed from the ground up with processing massive amounts of data in mind.

An emerging technology called phase-change memory (PCM) may be a part of that future design. This type of memory uses less power than conventional technologies, and it is also well suited to the development of architectures in which memory and processing are colocated, further reducing energy consumption and greatly enhancing processing speed. Moreover, many implementations of PCM are nonvolatile, which means that they can maintain their state after the power is turned off, enabling them to pull double duty as permanent storage.

But in reality, most PCM systems to date have had issues with high switching power and drifting resistance states which compromise the integrity of the stored data over time. With problems like these, the devices will never find any use outside of a research lab. But that may change in the near future as the result of some work recently done by a team led by researchers at Stanford University. They have developed a novel type of PCM that is fast, requires very little energy for operation, and exhibits high levels of stability over time.

The memory is composed of a material with the memorable name GST467. It is composed of four parts germanium, six parts antimony, and seven parts tellurium. The GST467 is sandwiched between layers of a few other ultra-thin materials in a layered superlattice structure. By utilizing GST467, the memory is endowed with very fast switching speeds, and the unique structure enables low-power switching and stability. In fact, this PCM can retain its memory state for more than a decade.

At present, the memory cells are 40 nanometers in diameter, which is about half the size of a coronavirus capsid. This is an impressive first step, but the team believes they can further shrink the cells with additional work. The memory operates at less than one volt, which is considerably less than competing technologies. And considering that these units have switching speeds of about 40 nanoseconds, they might prove to have practical applications in the future.

The researchers hope that their insights will lead to the industry-scale adoption of their phase-change materials and device architecture for high-speed, low-power storage.