1 of 15

Neuromorphic

Tactile Sensing

TAC

2 of 15

Motivation

SKIN

LOT OF SENSORS

SPARSE IN SPACE

SPARSE IN TIME

EVENT DRIVEN CODING

3 of 15

How we vision the complete architecture

4 of 15

Neural coding with piezoresistors?

SA

RA

5 of 15

Readout Stage: A Sensor Front End Designer’s Take

Match readout circuit to sensor resistance

Conductance behaves better than resistance�⇒ fix voltage, read current

Regulated cascode with scale-down current mirror

small current (1μA)

sensor current (0-26mA)

neuron current (0-1.5μA)

constant-ish sensor bias voltage

current mirror bias voltage

cascode bias voltage

0 kPa

200 kPa

long transistor

wide transistor

25 mA

1.4 μA

W 640x96 : 2

Mirror�L=400 nm (too short)

neuron input current depends on neuron input voltage�(0…1.6 V)

sensor current at constant-ish 448mV bias

Ludicrous ratio

Well-behaved�I_sensor vs. pressure

[same slide as before, with designer’s totally qualitative annotations: it’s way too early for hard and solid numbers]�This regulated cascode sensor interface is a proof-of-concept how to operate the resistive sensor at a realistic constant bias voltage.

The transistor M_reg measures the voltage across the resistor and regulates it by feedback through the cascode transistor M_casc.�The bias current for the regulated cascode transistor can be held fairly low. Proper dimensioning would require noise analysis.

At constant voltage across the resistor, its current is a relatively well-behaved function of pressure.�The current mirror has an unrealistically large division ratio (design: 30720, in simulation ca. 20000, varies with current). �It should probably be cascoded for a more stable division ratio.

6 of 15

Readout Stage

7 of 15

Readout Stage: Differential Degenerated NMOS

8 of 15

SA Fibers: Several neurons compared

Axon Hillock

Simplified DPI

Leaky IF

9 of 15

Axon Hillock Coding of Piezoresistive

10 of 15

DPI Neuron Coding of Piezoresistive

11 of 15

LIF neuron Coding of Piezoresistive

12 of 15

Benchmark of different neurons (simulation)

Performance	Axon-Hillock	DPI	LIF
Power (50 Kpa) (uW)	167	54	384
Size (um2)	220.8	67 + 500(Cmem)	356
Range (kPa)	5-50	10-50	20-60 (unoptimized)

13 of 15

FA Fibers: Delta Coding Circuit

14 of 15

The TAC Sensor Design Group

Ella Janotte

Ph.D. Student

EDPR, IIT Genova

Sahil Shah

Assistant Professor

University of Maryland

Michele Mastella

Ph.D. Student

BICS, Groningen University

Christoph Maier

Fly on the wall

Ivy League

Tübingen

Thanks to

Elisabetta Chicca,Chiara Bartolozzi, Elisa Donati, Benjamin Tee,

Neutouch EU grant 813713

15 of 15

Cargo Cult Neuromorphic Design: Bridge Current Mismatch

Typical sensor current: O(1mA) — Typical neuron branch current O(100pA)

Pointless to optimize neurons for power when the sensor draws 10⁷ more current�… but can be made to work in simulation with current divider (Good skywater tools exercise)