IBM has an obsession with the human brain. Well... maybe not an obsession, but experts at Big Blue believe (for some time, we must admit) that the future of computing lies in its design, structure, and operation. That’s how we arrive at a new announcement revealing the creation of “artificial neurons” based on phase-change technology. The best part is that this type of neurons can reach a scale similar to what we find in current processors without affecting their operation.
How IBM’s Artificial Neurons Work
Artificial neurons. Technically, devices that behave very similarly to what we see in a real neuron. We could discuss the difference until the Sun cools, but the point is that these artificial neurons, 500 to be precise, work as calculated. IBM Research’s work is not limited to imitating the functioning of a neuron, but also seeks to reproduce its overall structure. Thus we see elements equivalent to dendrites, axons, nuclei, membranes, and somas. IBM Research’s advance is in what would be the membrane. In a biological neuron, the membrane is a “lipid bilayer”, formed by two layers with lipid molecules. In very relaxed terms, the membrane works as resistance and capacitor. At first it resists conductance, but once it reaches a certain limit, it allows electricity to pass to other neurons.
In IBM's artificial neurons, the neuron is replaced by a small block of germanium, antimony, and tellurium (GeSbTe), which possesses phase-change properties. This allows it to pass with relative ease from amorphous phase (resistance) to crystalline phase (conduction) by applying a little heat. After a brief “rest period”, the GeSbTe “membrane” changes again from crystalline to amorphous, ready to repeat the process. Another very striking aspect of IBM's artificial neurons is that they demonstrate a “stochastic” behavior, that is, their results always carry a pinch of randomness, just as happens in real neurons. This is due to the slight differences between amorphous phases with each cycle, which by extension affect subsequent crystalline phases.
What This Means for Computing
Now we come to the best part: The phase change can extend over trillions of cycles (our billions in the Spanish system) , and IBM’s presentation speaks of a theoretical production at a scale of 14 nanometers. What does this mean? Parallelism, perhaps like we have never seen it. The possibility of placing billions of artificial neurons on a chip is one step closer... but the software must not fall asleep. After all, we will need code that is capable of using that kind of hardware.
