Daily Archives: May 1, 2016

Emulating Brain Functions with a Memristor Computer

Chip designs at the atomic level may require emulating the functioning of the human brain, while upholding Moore’s Law. In fact, this might be the only way forward. All forward-looking semiconductor design organizations such as Intel, HP, Samsung, and others know this and this has sparked an exponential interest in the study of memristors.

Among all known electronic components, memristors alone are capable of emulating the brain. It is common knowledge now that a human brain performs far better than the fastest supercomputer, while consuming only about 20 watts of power. Supercomputers cannot emulate, rather only simulate the brain, but they consume thousands of watts and are expensive to the tune of millions of dollars.

At Santa Fe, N.M., Knowm Inc. has expanded its portfolio by adding three different types of memristors. They offer a new high-density die-only option, along with all the raw data manufacturers will need to perform their characterization. Although another organization HP/Hynix is also trying to build commercial memristors, Knowm has beaten them to the market by diversifying its offering of memristors. Knowm is now offering three models with slow, medium, or fast hysteresis, based on the material they are made of – tungsten, tin, or chromium.

Regardless of the basic metal-ion used for its manufacture, all memristors work in the same way. The device consists of two electrodes, with a layer of metal located close to one of them. As a voltage is applied across the electrodes, metal ions move through the device towards the electrode with the lower potential. The device has a layer of amorphous chalcogenide material, which is also its active layer. Metal ions moving through this active layer form a conductive pathway between the electrodes. As the pathways spin through the active material layer, the device resistance drops. When the direction of the applied potential is reversed, the conductive channels dissolve and the device resistance increases.

This characteristic makes the memristor a bipolar device. The memristor also emulates the way neurons in the brain learn. Neurons send pulses through their synapses to strengthen them, equivalent to lowering the resistance, or they do not send pulses, thus causing the synapses to atrophy, equivalent to increasing their resistance.

IBM, together with DARPA or the Defense Advance Research Agency, is developing programs to simulate the process with digital computers. Their software will span the spectrum of accuracy in modeling the way synapses work. However, memristors are the only true emulators of the brain and its functions, so far. Some researchers are going to the extent of erecting scaffoldings for connecting memristor-based emulators to large-scale models. This is creating the need for components such as those Knowm is now offering.

Knowm’s offer is a treasure trove for researchers, being raw data from over 1000 experiments. It helps tremendously, as there is actually no well-defined specification to characterize the memristors properly. That allows researchers to generate their own characterization data, based on the properties of their choice from among the slow, medium, or fast hysteresis types. Knowm offers 16 memristors packaged in a ceramic single dual-in-line package. Their die only option holds 180 memristors.