Building a ‘brain’ computer with a powerful new memory
The Laboratory of Emerging Memory and Novel Computing (LEMON), led by Professors He Qian and Huaqiang Wu at Tsinghua University, has dedicated itself to working on memristor-based neuromorphic computing in the past few years. More recently, they demonstrated a fully hardware-implemented CIM system to efficiently realize the convolutional neural network (CNN), which is one of the most significant models for a neural network, for the first time.
Several members of the Tsinghua LEMON team standing by the memristor CIM system
With the rapid development of deep learning, artificial intelligence (AI) is able to achieve a better performance than human beings in many cognitive tasks, such as image recognition, natural language processing, and gaming. The growth of AI actually leads to higher and higher demands for smart chips with a high speed, low power, and high energy efficiency. As a comparison, AlphaGo, which is famous for beating the world Go champion, burns about 150,000 watts of power, while the brain of its human opponent consumes just 20 watts. In order to build more power-efficient AI hardware, researchers have been searching for better computing paradigms beyond the conventional von Neumann architecture, and new materials and devices to replace the mainstream silicon counterparts.
Among them, memristor, a new type of resistive switching memory, has been considered as one of the most promising technologies. It has many advantages such as high speed, low power, simple structure and small size. More interestingly, its ionic motion-based working mechanism resembles the biological synapses and neurons in our brain, and it can be made into high-density cross-point array to enable massively parallel multiply-accumulate computing inside memory (CIM) via physical laws. Therefore, it provides an appealing platform to overcome the so-called von Neumann bottleneck and boost computing efficiency, which has motivated researchers to build a powerful ‘brain’ computer using memristor.
Inspired by the human brain, memristor devices are organized into cross-point arrays to realize massively parallel in-memory computing and boost power efficiency.
The Laboratory of Emerging Memory and Novel Computing (LEMON), led by Professors He Qian and Huaqiang Wu at Tsinghua University, has dedicated itself to working on memristor-based neuromorphic computing in the past few years. The group has made tremendous achievements from material and device engineering, process development to circuit and chip design, as well as algorithm and system demonstrations. Working with a commercial silicon foundry, the team have developed a hybrid integration route to fabricate large-array memristors with optimized material stacks as a flexible hardware test platform. More recently, they demonstrated a fully hardware-implemented CIM system to efficiently realize the convolutional neural network (CNN), which is one of the most significant models for a neural network, for the first time. This breakthrough was published last month in Nature journal entitled “Fully hardware-implemented memristor convolutional neural network”. With the innovative hybrid training and spatial parallel computing techniques developed in this work, the fully hardware-based memristor CNN achieved an energy efficiency more than two orders of magnitude greater than that of graphics-processing units.
Photo of the CIM system with multiple memristor arrays to implement CNN.
This month in ISSCC, the top conference in the field of integrated circuits, the team reported the first complete CIM chip to implement a multi-layer percepton neural network for the classification of handwritten digits from the MNIST dataset. The chip integrated nearly 160,000 memristors together with all the peripheral circuits on chip, and achieved an ultra-high energy efficiency of 78.4 tera operations per second per watt. The running power is as low as 40 milli-watts and the recognition accuracy is up to about 95% for classifying MNIST images. The team is now working on building more complex CIM chips and scaling up the memristor array size to further boost the system performance, taking advantage of memristor.
Photo of the MNIST chip and demo system for handwritten digit classification.
Looking forward, there still a long way to go to build a ‘brain’ computer with memristor as there are still many challenges ahead in both hardware and software. The group has been working closely with researchers across different departments in Tsinghua and also leading experts worldwide. They believe that such interdisciplinary research and collaboration are critical to break through traditional thinking and establish a completely different computing system from existing ones, which will hopefully revolutionize AI hardware with this powerful memristor.
（Source: The Laboratory of Emerging Memory and Novel Computing）
