Alibaba Unveils 5nm CPU; Introducing Self-backpropagation; Peking U Announces School of Computer
China’s AI news in the week of October 24, 2021
Alibaba Introduces World’s First 5nm Arm-Based Server CPU
The world’s largest technology companies like Google, Amazon, and Facebook and their Chinese tech counterparts are rushing to develop their semiconductors to suffice their increasing computation demands and reduce their billion-dollar budgets on purchasing third-party chips from Intel and AMD.
This week, Alibaba Cloud, the e-commerce giant’s cloud and tech unit, has joined the gang with a home-grown offering: a 5nm ARM-based CPU for cloud servers. Named Yitian 710 (倚天, a powerful weapon originated from famous Chinese novel The Heaven Sword and Dragon Saber), the chip features:
- 128 Arm cores with 3.2GHz top clock speed (the largest number of cores on a single chip as far as we know)
- Each processor chip has 60 billion integrated transistors.
- The first server processor compatible with the latest Armv9 architecture
- 8 DDR5 channels and 96-lane PCIe 5.0
- 440 in SPECint2017 (a standard benchmark to measure CPU integer processing power), surpassing that of the current state-of-the-art Arm server processor by 20% in performance and 50% in energy efficiency.
How good is Yitian 710? Below is a quick comparison between Yitian 710 and other top server chips (illustrated by a Chinese semi-expert).
Alibaba is urged to custom chips to support their gigantic e-commerce platform and digital payment systems. For example, Alibaba’s Alipay system processes six hundreds of thousands of transactions per second with a latency of no more than 20ms.
Yitian 710 is the newest addition to China’s push for semiconductor independence against the backdrop of the U.S.’ increasing intent of decoupling from China. However, despite its superior features, Yitian 710 may still hit the same roadblock against Huawei. The chip is built on the Arm v9 architecture and will be possibly mass-produced by Taiwan Semiconductor Manufacturing Co. Arm and TSM could stop licensing their chip architecture and cut supplies respectively to comply with the U.S. export control as it did to Huawei. However, the scenario will be implausible shortly, as Alibaba uses its chip only for its cloud servers. “We plan to use the chips to support current and future businesses across the Alibaba Group ecosystem. We will also offer our clients next-generation computing services powered by the new chip-powered servers in the near future,” said Jeff Zhang, President of Alibaba Cloud Intelligence and Head of Alibaba DAMO Academy.
Alibaba Cloud also announced they would open-source the XuanTie IP core series code, Alibaba’s custom-built processors based on RISC-V instruction-set architecture.
Institute of Automation Introduces Self-backpropagation
Widely used for training deep neural networks, backpropagation (BP), short for “backward propagation of errors,” is a classic algorithm by tuning weights of each node to achieve the correct output. However, its significant computational cost remains a hurdle to churn out compute-efficient AI applications.
Recently a team of researchers from the Institute of Automation, Chinese Academy of Sciences, applied a concept discovered from biological neural networks to train neural networks that can significantly reduce the computational cost while increasing accuracy compared to previous methods. The paper Self-backpropagation of synaptic modifications elevates the efficiency of spiking and artificial neural networks has been published on Science Advances.
The idea of self-backpropagation (SBP) could be originated in a 1997 paper, which studied that hippocampal neurons, a well-known seahorse-shaped region of the brain that is responsible for learning, emotions, and memory, can spread synaptic plasticity like long-term depression (LTD) from output synapses to input synapses. Such spread occurs in three forms: Presynaptic lateral spread, Postsynaptic lateral spread, and Backpropagation. The research team focused on SBP as the SBP phenomenon is also found in developing retinotectal circuits as well.
The phenomenon of SBP represents a form of nonlocal activity-dependent synaptic plasticity that may endow developing neural circuits the capacity to modify the weights of input synapses on a neuron in accordance with the status of its output synapses.
Researchers then examined the effects of applying SBP on the accuracy and computational cost of SNNs and ANNs (the restricted Boltzmann Machine Network) for three different benchmark tasks — recognition of handwritten digits, phonetic transcription, and gesture recognition.
Overall, experiments showed that SBP improved the efficiency of SNNs by increasing the accuracy (up to 1.4%) and reducing the computational cost (up to 79.6%) in three benchmark tests. The same benefits were demonstrated in training ANN-BMNs as well as the introduction of SBP reduced both the error rate (up to 4.2%) and computational cost (up to 74.3%) in three benchmark tests.
Peking University to Establish School of Computer
China’s prestigious Peking University recently announced it would establish School of Computer (Science), School of Electronics, and School of (Machine) Intelligence to renovate its engineering and information science education. The news was released at the university’s first-ever New Engineering Education International Forum 2021.
The new School of Computer is a structural upgrade of the pre-existing Department of Computer Science and Technology. It will encompass seven sub-institutes and centers such as the Institute of Computational Linguistics and the Institute of Digital Media. The school has 118 faculties responsible for cultivating 1,489 full-time students in four bachelor majors and two graduate majors. Yang Fuqing, a well-respected professor at the School of Information Science and Technology and a member of the Chinese Academy of Sciences, will serve as the Honorary Dean of the school.
Peking University is determined to stand at the forefront of China’s New Engineering Education (NEE) reform, a nationwide initiative to cultivate innovative talents in applied engineering science and technology. Professor Tengteng Zhuang from the Chinese University of Hong Kong described NEE in his 2018 paper that:
China has pushed forward the NEE reform with measures such as formulating National Standards for dozens of categories of engineering programs, commissioning 600+ research projects on NEE development, establishing new engineering programs and interdisciplinary courses, strengthening university-partnership, updating accreditation for engineering programs, and improving both external and internal quality assurance mechanism.
Over the past few years, Peking University has set up the Academy of Artificial Intelligence, College of Future Technology, Department of Materials Science and Engineering, and School of Integrated Circuits.
- Xpeng’s flying car company, HT Aero, has raised over $500 million in its Series A funding round, led by IDG Capital, 5Y Capital, and XPeng with participation by a consortium of renowned investors, including Sequoia China, Eastern Bell Capital, GGV Capital, GL Ventures and Yunfeng Capital. Founded in 2013, the Tianhei, Hubei-based urban air mobility company is planning to roll out a flying car in 2024.
- Dreame Technology, an innovative consumer product company that focuses on smart home cleaning appliances, has raised RMB 3.6 billion yuan in its Series D funding round. Known as the Chinese Dyson challenger, the Suzhou-based company will invest the raised capital in high-speed digital motor and artificial intelligence.
- JIMU Intelligent, an intelligent driving solution provider, has bagged RMB 200 million yuan in its Series C1 funding round led by Forebright Capital and SDIC Unity Capital. Founded in 2011, the Wuhan-based company aims to provide global OEMs and Tier-1 suppliers with self-developed leading intelligent driving solutions based on multi-sensor fusion and domain controllers.