2.4. Measurement setup and characterization methods
As detailed in Section 2.1, the entire system is ratiometric with respect to the clock frequency. That is, the system clock can be scaled so that the neuromorphic system operates anywhere from biological real time up to a factor 100 faster. As operation at biological real time is the most challenging in circuit terms as well as the most interesting in terms of computation, real-time operation was used for the measurements in this paper. The corresponding clock frequency is 3.3 MHz, generated by a configurable clock divider from the 330 MHz central system clock. At this frequency, the synaptic matrix update period is 0.62 ms (compare Section 2.1).
As the different leakage currents of MOS switches are highly temperature dependent, we investigated how well our low-leakage switch technique operates at different temperatures. Thus, the measurements of the presynaptic adaptation are carried out at the temperatures indicated by using the temperature controlled setup shown in Figure 9. The IC package is held at the adjusted temperature with ca. ±2 °C deviation. The output of the presynaptic adaptation can be measured either via tracing the PSC time course from one of the analog test outputs or indirectly by monitoring the spike output of a connected neuron. Directly measuring the PSC voltage via an oscilloscope is well-suited for detailed short-time measurements, which we used to verify correct operation of the circuitry. For reducing noise in this case, the aquired waveform data was averaged over time bins of 0.1–0.3 ms.
Figure 9  Setup for measurements with controlled temperature. Direct oscilloscope measurements are less practical for automatic extraction of a multitude of time constants. For this case, we used the following purely spike-based protocol: The adaptation state is probed by sending an input spike and counting the number of output spikes in reaction. For getting a reasonably strong response, the synaptic weight and the PSC scaling voltage are set to their maximum values. Setting the membrane time constant to a high value as well, the number of output spikes per input spike is approximately linearly dependent on the PSC amplitude. For the measurements, we only activated depression, so that the PSC amplitude of a spike directly resembles the current state of the depression variable. For each time constant measurement, the depression variable is charged by initially applying 10 spikes. Afterwards, the adaptation strength is set to zero, so that the depression variable relaxes back to its resting state. This relaxation is monitored by continuously probing the state with input spikes. From the relaxation time course, the time constant is extracted by calculating the best-fitting (smallest root mean squared error) exponential function, with amplitude and time constant as free parameters. Results are averaged over 10 repetitions.
The measurements of the stop learning synapses are carried out at ambient temperature, i.e., no special measures for chip cooling are taken.