At the bottom of the page is a video showing the key principle in action.

Neuromorphic processing elements, often using individual FET devices, can create very low power analog signal processing elements. They leverage modern processes abilities to make high density connections and small devices. The circuits work better as the processes shrink  – thus these analog techniques are a good fit to sub 10nm processes. They represent a new way to do analog design: functionality through complexity of interconnection.

Tracking a Signal Envelope

Using a complex interconnection of thousands of tiny devices results in a very low power circuit that accurately tracks the envelope of a signal. The resulting circuit can be shown as a pair of NMOS devices, but in fact each of the NMOS contains within it 512 small NMOS devices.

We will find that none of the advantages of analog design are ignored: many thousands of devices have low noise and good matching when working as aggregate connections of many devices.

The new fact is the complex way in which connections between the tiny devices are made. This is an example.

1 – Two NMOS devices, each is a sub-circuit of 512 smaller devices connected in a complex pattern.

If a multi-phase sinusoidal signal is applied to the four input wires on the left we would expect that the common source would track the peak: the action of an NMOS device is to turn on further as the gate voltage rises – we have a set of source-followers and the highest gate voltage dominates.

That is indeed the case in the U2 device on the right. It has 512 devices within it and all are connected in series/parallel.

But the device U1 on the left also has 512 devices within it connected differently. The connection scheme results in the U1 device source tracking the minimum of the signal.

It is far from evident how NMOS devices can do that: it is a phenomenon that emerges due to the complexity of interconnection.

The power consumption is as shown – just 20nA  – very low power.

(We at SiliconIntervention use this circuit to track the envelope of the complex signal of the cepstrum of the audio signal  – in other words to calculate the MEL bin spectrum in our analog FFT and MFC calculator. It is a small example of the processing power based on very complex connections of thousands of tiny devices. It is an example of the new analog).

The Principle of Operation

The video (below) shows the step by step approach to making the neuromorphic example.

If you are familiar with the concept of “stacked devices” as used in Cadence you will note the similarity. However, SiliconIntervention adds one key extra feature: we add a deliberate systematic perturbation to the device interconnection which gives rise to the additional features.

In the usual use of stacked devices a mean-of-means improvement of noise and offset is augmented with known common centroid techniques to reduce variability and allow very small geometries to have reasonable matching.

The innovative step in conventionally stacked devices is the use of multiple small elements as opposed to the creation of larger devices in the less advanced processes. This is proven to be the optimum way to get good analog matching on very small (<5nm) processes.

We at SiliconIntervention recognize a further possibility: do not connect all the small devices in simply series or parallel, rather add a systematic variability in connectivity which causes the mean-of-means statistical improvement to coalesce around an emergent property.

What does “coalesce around an emergent property” mean?

The best example is probably the averaging of errors in a flash convertor (see for example, B Murmann, EE315B – Chapter 7, Slide 16/17 of Standford EE315B VLSI Data Conversion Circuits) where we note that the averaging effect is to more closely approximate a line rather than a point: connectivity from adjacent comparators is not improving the offset of any given comparator, rather it is improving the fit of all the comparators to a straight line.

SiliconIntervention’s technique is similar, but is not restricted to a line or point: more complex effects of the aggregate behavior of many small devices can be achieved. This circuit that finds the minimum and the maximum of a signal is an example.

  • This example is one of static aggregate behavior: you will note that the connectivity and configuration is fixed and does not evolve with time. Other more powerful examples exhibit dynamic aggregate behavior where time-varying emergent properties can be created.

The video below shows an example of behavior of an analog circuit that depends upon the complexity of its interconnection.