IBM Demos In-Memory Massively Parallel Computing

　　Today’s experimental non-von Neumann computing architectures mostly make use of memristive devices modeled on the human brain; they do not separate data memory from computing hardware and thus avoid the inefficiency of von Neumann computers’ repeated load/store operations. Now IBM Research (Zurich) has demonstrated a way to mass-produce 3-D stacks of phase-change memory (PCM) to perform memristive calculations 200 times faster than von Neumann computers. The in-memory coprocessor uses algorithms that exploit the dynamic physics of phase-change memories simultaneously on myriad cells, similar to the way millions of neurons and trillions of synapses in the brain operate in parallel.

　　The development, which IBM will demonstrate in December at the International Electronic Devices Meeting (IEDM), could return the company to the brink of hardware dominance.

　　“We have demonstrated that computational primitives using non-von Neumann processors can be used to do machine learning tasks,” IBM Fellow Evangelos Eleftheriou told EE Times. “So far, we predict a speedup of 200 times for our non-von Neumann correlation detection algorithm compared to using state-of-the-art computing systems, but we have many other computational primitives on the way that we will demonstrate later this year.”

　　The new paradigm combines PCM crystallization dynamics with an acceleration methodology called in-memory computing, which loads all data into RAM instead of swapping data sets into and out of mass memory (hard drives or flash). IBM’s approach does not force the in-memory values through the von Neumann bottleneck of a central processing unit; rather, it leaves the initial-state memory values in each PCM cell and uses a specialized memory controller to perform simultaneous, parallel operations on the cells. Calculations are performed in place by harnessing the physical properties of the phase-change RAMs.

　　Building on crystallization-dynamics discoveries

　　Memristive non-von Neumann architectures work like the brain by strengthening (lowering the resistance between) memory synapses each time they are used (and, conversely, increasing the resistance over time if they are not frequently used). The pattern-recognition and other algorithms get increasingly accurate as they gain experience, “memorizing” similar data sets and “forgetting” irrelevant ones whose patterns are seldom repeated. IBM uses this technique in its all-digital neurocomputer e-brains, which are already shepherding the U.S. nuclear arsenal and piloting U.S. fighter jets.

　　IBM Zurich’s latest effort does not emulate brainlike algorithms such as the digital spiking used in its neurocomputer e-brains. Rather, the development builds on IBM’s discoveries about the crystallization dynamics of phase-change memories.

　　“What we are trying to do is make more energy-efficient processors by avoiding all the load/store operations” of a von Neumann computer, IBM Research Staff Member Abu Sebastian said in an interview. “Today we’re showing how to use crystallization dynamics to perform unsupervised deep learning, but eventually we plan to build a coprocessor that will allow a von Neumann computer to offload all sorts of tasks it is ill-suited to perform well.”

　　The prototype houses 1 million in-memory cells, each performing the same deep-learning computational tasks on the unique data set loaded into it. The memristor-like use of PCM crystallization dynamics both accelerates time-to-results and eliminates power-wasting load/stores. IBM says the technology should be easily scalable both horizontally and vertically to realize 3-D non-von Neumann coprocessors that can solve tasks of almost any size.

　　In more detail, the PCM device uses a germanium-antimony-telluride alloy, sandwiched between two electrodes. When pulsed, the phase-change material shifts from an amorphous to a crystalline phase in easily controllable resistance steps that vary from extremely low (for 0) to extremely high (for 1) or anywhere in between (analog operation).

　　Model of the phase-change material.

　　Sebastian was the lead author on a paper describing the development in Nature Communications. He also leads a European Research Council project on the same topic.