Periodic Reporting for period 2 - QUEFORMAL (Quantum Engineering for Machine Learning)
Periodo di rendicontazione: 2020-01-01 al 2022-12-31
The present reality of artificial intelligence (AI) applications is that algorithms run on huge data centers, consuming large amounts of energy. This reality is not adequate to respond to an emerging need: more and more applications requiring video and audio recognition, autonomous vehicles and robots, would benefit from cognitive functions performed locally (in ‘embedded systems’, technically), in order to reduce latency in cases in which delays are critical and in scenarios where a fast data connection is not available.
It has been recently shown that processor performance and memory energy efficiency cannot keep up with the exponentially increasing size and computational capacity required in deep neural networks (DNNs), one of most successful techniques for artificial intelligence. To win such challenges, significant improvements are required in device technology, architecture, and algorithms.
A possible solution pursued in the hardware AI community is moving to near-data processing architectures, in which computation is distributed in an array of Processing Elements, each containing both logic and memory, through which data flow to be processed. Use of embedded non-volatile memory is also pursued in order to further reduce energy consumption.
The vision proposed within the QUEFORMAL project surpasses both the present paradigm and the paradigms under development: Quantum engineering of transistors based on lateral and vertical heterostructures of 2D materials unifies materials science and device engineering exploiting the degrees of freedom provided by such materials to fabricate – on the same platform and in close vicinity – both low-voltage transistors and non-volatile memories.
The overall objective and targeted breakthrough of QUEFORMAL is to demonstrate the fabrication and operation of devices for logic-in-memory integrated circuits based on vertical and lateral heterostructures of 2DMs and to show the potential of this technology for integrated circuits for embedded machine learning capabilities.
QUEFORMAL pursues a very risky and original target, with the extremely high potential gain of advancing a science-enabled technology for the fabrication of integrated circuits for machine learning, in a field in which Europe has strong basic-science leadership, thanks to the pioneering breakthroughs on graphene and 2D materials.
It has been recently shown that processor performance and memory energy efficiency cannot keep up with the exponentially increasing size and computational capacity required in deep neural networks (DNNs), one of most successful techniques for artificial intelligence. To win such challenges, significant improvements are required in device technology, architecture, and algorithms.
A possible solution pursued in the hardware AI community is moving to near-data processing architectures, in which computation is distributed in an array of Processing Elements, each containing both logic and memory, through which data flow to be processed. Use of embedded non-volatile memory is also pursued in order to further reduce energy consumption.
The vision proposed within the QUEFORMAL project surpasses both the present paradigm and the paradigms under development: Quantum engineering of transistors based on lateral and vertical heterostructures of 2D materials unifies materials science and device engineering exploiting the degrees of freedom provided by such materials to fabricate – on the same platform and in close vicinity – both low-voltage transistors and non-volatile memories.
The overall objective and targeted breakthrough of QUEFORMAL is to demonstrate the fabrication and operation of devices for logic-in-memory integrated circuits based on vertical and lateral heterostructures of 2DMs and to show the potential of this technology for integrated circuits for embedded machine learning capabilities.
QUEFORMAL pursues a very risky and original target, with the extremely high potential gain of advancing a science-enabled technology for the fabrication of integrated circuits for machine learning, in a field in which Europe has strong basic-science leadership, thanks to the pioneering breakthroughs on graphene and 2D materials.
The main results obtained by the project in the four years are the following:
- Design, fabrication and experimental demonstration of a neural network obtained with devices based on two-dimensional materials such as molybdenum disulphide and heterostructures of molybdenum disulphide and niobium disulphide.
- Multiscale modeling and system-level analysis of the performance of a large analog neural network based on 2D materials.
- Design, fabrication and experimental demonstration of an analog neural network in silicon CMOS technology
- Fabrication of improved contacts to two-dimensional materials based on lateral or vertical heterostructures
- Improvement of two-dimensional materials growth techniques (for PtSe2, PtS2, NbS2, MoS2)
- Expansion and extension of a multi-scale simulation framework for materials, devices, circuits and systems based on 2D heterostructures
- Design, fabrication and experimental demonstration of a neural network obtained with devices based on two-dimensional materials such as molybdenum disulphide and heterostructures of molybdenum disulphide and niobium disulphide.
- Multiscale modeling and system-level analysis of the performance of a large analog neural network based on 2D materials.
- Design, fabrication and experimental demonstration of an analog neural network in silicon CMOS technology
- Fabrication of improved contacts to two-dimensional materials based on lateral or vertical heterostructures
- Improvement of two-dimensional materials growth techniques (for PtSe2, PtS2, NbS2, MoS2)
- Expansion and extension of a multi-scale simulation framework for materials, devices, circuits and systems based on 2D heterostructures
The project has obtained a few first results, that represent a significant improvement with respect to the state of the art:
- First neural network fabricated with 2D materials
- First Analog non volatile memory fabricated with 2D materials
- First experimental demonstration of retention and temperature resilience of analog neural network fabricated in silicon CMOS technology
- First experimental demonstration of p-doped contacts on molybdenum disulphide
A clear impact of the project is that it clearly showed a promising path in the vertical integration of 2D materials with BEOL processing over a CMOS circuit.
- First neural network fabricated with 2D materials
- First Analog non volatile memory fabricated with 2D materials
- First experimental demonstration of retention and temperature resilience of analog neural network fabricated in silicon CMOS technology
- First experimental demonstration of p-doped contacts on molybdenum disulphide
A clear impact of the project is that it clearly showed a promising path in the vertical integration of 2D materials with BEOL processing over a CMOS circuit.