13
NEUral computing aRchitectures in Advanced Monolithic 3D-VLSI nano-technologies EFECS Lisbon 21 November 2018

NEUral computing aRchitectures in Advanced …in Advanced Monolithic 3D-VLSI nano-technologies EFECS Lisbon 21 November 2018 Project Objectives: 1. Develop ultra-low power neuromorphic

  • Upload
    others

  • View
    10

  • Download
    0

Embed Size (px)

Citation preview

NEUral computing aRchitecturesin Advanced Monolithic 3D-VLSI

nano-technologies

EFECS

Lisbon 21 November 2018

Project Objectives:

1. Develop ultra-low power neuromorphicarchitectures for embedded ArtificialIntelligence

2. Implement in advances technologies andexploiting large TFT-based inter-chiprouting, FDSOI power efficient electronics,new Resistive Memories and Monolithic3D integration

3. Investigate specific machine learningalgorithms to co-optimize the algorithms-architecture-devices systems

• Project Structure:

Project Expected Impact:

• Provide a European scalable embedded solution (material and development environment) for distributed Artificial Intelligence at the Edge

• Mixed signal spiking neural network chip in ST FDSOI 28nm technology: Architecture

• Mixed signal spiking neural network chip in ST FDSOI 28nm technology:

DynapSEL chip and boards

• Mixed signal spiking neural network chip in ST FDSOI 28nm technology: Benchmark

• Common test vehicle for RRAM and CMOS tests

• RRAM as analog or digital synapses, new materials

• Area is 1,8mm². It contains 10 neurons and 1440 synapses, (11,5k OxRAMs)

• Circuit 1: Area is 0.2mm², 4k OxRAMs (binary synapses), 64 pre-neurons, 64 post-neurons

• RRAM based TCAM and spiking neural network

• Indium Gallium Zinc Oxide TFT based BEOL switching matrix

Measurement of X[0] to Y[2} pulse propagation in segmented bus TFT chip

Low thermal budget steps (here epitaxy demonstration) to achieve top leayr device performances without degrading the bottom

RRAM integrated in the top layer but connected to both layers (ltop) and in the bottom layer Front End (bottom)

• Monolitich 3D integration

• Benchmark and applicationsTask: detect and classify anomalous beats online• important: small number of false negatives (missing anomalies)• important: small number of false positives (false alarms)• difficult even for human experts

Method:• Train reservoir computing (RC)

neural network• Structure biologically inspired (mix

of excitatory and inhibitory neurons)

• First, optimize using digitally simulated, high-precision reservoir baseline

• Second, transfer to low-precision, spiking reservoir

• Compare with internal baseline and published previous results

Results / status:• Spiking nets still simulated on

BRIAN• No pre-processing• Digital/spiking accuracy .91/.88,

sensitivity .69/.63, precision .98 /.88

• Symmetric scores: accuracy .91/.88, sensitivity .84/.80, precision .90/.88

• Benchmark and applicationsExperimental setup:• Previous generation (180nm Dynap circuit (University of

Zurich, ERC neuroP Grant 257219) • Dedicated PFGA card and segmented bus TFT IC by IMEC• Liquid State machine trained offline and mapped on Dynap

box.

ECG heart beat estimation & classification• We are currently solving ECG

classification with analog VLSI neurons• We are using 320 neurons (256+64), as a

reservoir. • An ECG complex creates about 200/300

spikes in the reservoir.• The energy cost per spike is about 2.9pJ,

therefore detecting a peak costs about 870 pJ of energy.

• At an average of 60 beats per minutes we will consume about ~ 15 nW for ECG beat estimation (4µW/MHz of power on ARM Cortex-M0 state of the art).

• Conclusions

• The project has led to the fabrication of the most efficient spiking network in the word

• Technology building blocks for further gains and improved integration have been developed

• Dedicated network and algorithms for embedded AI have been validated and will be demonstrated in real life application by the end of the project