Phase change memory can operate thousands of times faster than current RAM
by Joel Hruska
As existing scaling methods for memory and logic (CPUs) come to an end, researchers have been searching for next-generation alternatives to conventional DRAM and circuit design. Phase change memory has a lot to recommend it as a long-term replacement for DRAM — it’s fast, it can retain data for long periods of time, and Intel’s Optane / 3DXPoint is thought to be partially based on phase change memory technology.
One of the principle concerns with any new technology is how much the new design can potentially scale. Switching to a new memory type isn’t trivial and chip designers want to know what kind of performance they can expect in 5–10 years, not just the next 18–24 months. Now, new research from a 19-member team of scientists led by Aaron Lindenberg at Stanford has shown that phase change memory can begin to switch between its two states (amorphous and crystallized) on the picosecond timescale. Conventional DRAM operates at nanosecond timescales, which means phase change memory could theoretically be thousands of times faster than conventional RAM.
In theory, phase change memory could eventually present a solution to the so-called memory wall, or memory gap. The major problem with DRAM is that while memory clock speeds have increased enormously over the past 30 years, memory latency has dropped much more slowly. Typically latency increases with every clock speed jump; DDR4–3200’s memory cell cycle time in nanoseconds is roughly on par with DDR3–1600, and DDR3–1600 is on-par with DDR-400.
Phase change memory (PCM) works by shifting between two states: An amorphous state with high resistance and a crystalline state with low resistance. The question that Dr. Zalden’s team set out to answer is how quickly does phase change memory begin to change from one to the other, and whether we can take advantage of the speed.
Here’s where things get a bit odd. What Dr. Zalden’s team found is essentially this: Exposing phase change memory to a 0.5THz electric pulse for picoseconds at a time can create crystallized filaments that can be measured and theoretically used to store data. Remember, memory fundamentally works by creating a measurable state we can label as a 0 or a 1 (on or off). The ability to create crystal filaments within a memory cell at the picosecond timescale means that next-generation memory might finally close the latency gap with CPUs.
These crystalline filaments come into existence while the bulk of the material is still amorphous — but they represent a difference in state that can be measured and therefore used. If you need a rough analogy, imagine a marsh or swamp that stretches for miles. People who wish to cross the swamp must trek across it on foot. Draining and dredging the swamp would be extremely expensive and time-consuming. What this research paper shows is that it’s possible to create a floating bridge across the swamp without going to the trouble of draining the entire area — and this bridge can also be reversed (removed, in our analogy) without harm to the underlying material. If you leave the bridge in place, it’ll stay there and continue to provide a useful function — drastically reducing how much time it takes to cross from Point A to B.
That said, there are some significant barriers between us and memory that’s orders of magnitude faster than anything on store shelves today. THz pulses haven’t been created in circuit boards yet as far as we’re aware; the signal doesn’t propagate well down the thin copper wires used in modern designs. Now that we know that phase change memory can switch so quickly, however, we also know that there’s more potential headroom in design than some may have previously thought.
This is more critical than you might think. One issue we’ve covered repeatedly at ExtremeTech is the rapidly rising cost of next-generation fabs and foundries. These costs have to be justified by product sales and overall performance — and while you might think this makes companies hungry to find solutions, it also makes them wary of betting on unproven technologies that might never fulfill their potential. Proving that phase change memory can operate at picosecond timescales doesn’t guarantee that the tech will replace DRAM, but it does demonstrate that PCM could potentially operate at frequencies that DRAM can’t touch.
We spoke to Peter Zalden, first author on the paper, who confirmed to us that the strength of today’s research is that it confirms the speed of PCM while simultaneously offering a much lower-power solution due to the absence of leakage currents. Given that Intel’s Optane memory technology is at least theorized to rely on some principles of phase change memory it’s possible that we’ll see a renewed focus on this technology over the next several years.
Now read: How do SSDs work?
Originally published at www.extremetech.com.