Printed Computing: Enabling Extremely Low Cost Pervasive Near Sensor Computing

Projektinformationen

PRICOM

ID Finanzhilfevereinbarung: 101052764

DOI

10.3030/101052764

EK-Unterschriftsdatum 29 Juni 2022

Startdatum 1 Oktober 2022

Enddatum 30 September 2027

Finanziert unter

• European Research Council (ERC)

Gesamtkosten

€ 2 499 286,25

EU-Beitrag

€ 2 499 286,00

Koordiniert durch

KARLSRUHER INSTITUT FUER TECHNOLOGIE
Germany

Periodic Reporting for period 1 - PRICOM (Printed Computing: Enabling Extremely Low Cost Pervasive Near Sensor Computing)

Berichtszeitraum: 2022-10-01 bis 2025-03-31

Zusammenfassung vom Kontext und den Gesamtzielen des Projekts

Computing has yet to penetrate cost-sensitive domains such as fast-moving consumer goods and personalized medicine, largely due to the high production costs of rigid silicon technologies. Additive manufacturing for printed electronics offers a low-cost, flexible alternative, but conventional digital architectures are unsuitable due to limited device counts, large feature sizes, and high variability in printed components.

The PRICOM project develops unconventional mixed-signal classifier paradigms optimized for additive printing and flexible electronics, aiming to minimize hardware footprint while directly processing analog sensory inputs with high classification accuracy. Key contributions include:

1. Printed Mixed-Signal Classifiers – Novel architectures and tailored training algorithms that process analog sensory data with far smaller area than digital counterparts, maintaining accuracy.

2. Design Space Exploration – Hardware-in-the-loop optimization of printed ML primitives, balancing accuracy, reliability, area, energy, and performance.

3. Reliable Physical Design – Data-driven layout rules and variation-aware algorithms to improve yield and reliability under manufacturing variability.

4. In-Situ Post-Fabrication Tuning – Iterative tuning to correct imperfections and enable point-of-use customization.

5.Analog Function Approximation – First framework using analog building blocks for compact, low-power computation in flexible electronics.

PRICOM advances computing from materials and devices to complete systems, enabling low-cost, pervasive computing in emerging domains.

Arbeit, die ab Beginn des Projekts bis zum Ende des durch den Bericht erfassten Berichtszeitraums geleistet wurde, und die wichtigsten bis dahin erzielten Ergebnisse

The PRICOM project has delivered multiple breakthroughs in printed and flexible electronics for machine learning and neuromorphic computing, advancing beyond the state-of-the-art:

1. Analog Printed Spiking Neuromorphic Circuits – First fully functional spiking neuromorphic system using inorganic printed electrolyte-gated transistors (EGTs), enabling energy-efficient processing on flexible substrates.

2. Cost-Effective On-Sensor Learning – Reduced ADC front-end cost and complexity for training printed MLPs, improving scalability and resource efficiency.

3. Approximate Computing for Printed Circuits – Co-design methodology for approximate MLPs that optimizes power, area, and reliability while preserving high accuracy.

4. Model-to-Circuit Cross-Approximation – Automated framework combining algorithmic approximation, logic pruning, and voltage over-scaling to cut area and power without major accuracy loss.

5. Bespoke Robust Printed Neuromorphic Circuits – Custom designs with device-level optimizations for stable operation under manufacturing variability.

6. SpikeSynth – Adaptive analog printed spiking neural network framework achieving robust, low-power computation through device-level adaptability.

7. PRINT-SAFE – Ultra-low-cost printed electronic design with scalable adaptive fault endurance for resilience under variations and stress.

8. Automatic Test Pattern Generation – Framework for efficient fault detection in printed neuromorphic circuits, tailored to variability constraints.

9. Low-Power Flexible Stress Monitoring – Energy-efficient classifiers for continuous wearable stress monitoring on flexible substrates.

10. Binary Search ADC for Flexible Classifiers – First co-designed binary search ADC with in-training calibration to improve accuracy and efficiency under device variability.

11. Analog Function Approximation – First framework using analog building blocks in flexible electronics for compact, low-power computation.

These innovations establish a new paradigm in printed computing, addressing energy efficiency, cost, reliability, and ML optimization for low-cost, flexible, and power-constrained applications.

Fortschritte, die über den aktuellen Stand der Technik hinausgehen und voraussichtliche potenzielle Auswirkungen (einschließlich der bis dato erzielten sozioökonomischen Auswirkungen und weiter gefassten gesellschaftlichen Auswirkungen des Projekts)

Printed Spiking Neural Networks
Biologically inspired Spiking Neural Networks (SNNs) offer energy-efficient neuromorphic computing for emerging domains such as soft robotics, wearables, and IoT. Printed electronics, leveraging soft materials and flexible substrates, provides a low-cost alternative to silicon.

We present the first complete spiking neuromorphic computing system implemented using inorganic printed electronics, featuring:

1. Programmable Energy-Efficient Spiking Neuron – N-type electrolyte-gated transistor (EGT)-based neuron optimized for low-voltage, energy-harvested edge applications.
2. Transformer-Based Learning Model – Differential Transformer architecture for task-specific training, ensuring adaptability and accuracy.
3. Performance Validation – Simulation-based validation and benchmarking on standard datasets, confirming efficiency and applicability.
4. Variation aware modeling - Robustness aware training of proposed SNN

This interdisciplinary approach merges neuromorphic computing, printed electronics, and ML to enable ultra-low-cost, energy-efficient systems for resource-constrained environments.

Model-to-Circuit Cross-Approximation for Printed Classifiers
We introduce the first automated cross-layer approximation framework for printed ML classifiers (MLPs, SVMs), combining:
a) Algorithm-level coefficient approximation,
b)Logic-level netlist pruning,
c) Circuit-level voltage over-scaling.

Evaluations on 12 MLPs and 12 SVMs (6000+ designs) show 51% area and 66% power reduction with <5% accuracy loss. Notably, 80% of classifiers run on battery power while maintaining near-identical accuracy to exact designs.
Key contributions:
1. First Holistic Cross-Approximation Evaluation for printed ML classifier design.
2. Automated Optimal Design Generation under battery constraints.
3.Battery-Powered Approximate Computing enabling complex, low-power printed ML classifiers.