Discover the Rising Stars №012 | Interviews TensorChip Vice President Shang Huibin
Alibaba Cloud Innovation Center has introduced a new column, Discover the Rising Stars, to shed light on the backgrounds of entrepreneurs and company innovations. Through interviews, live streams, and other media approaches, we will deliver valuable news stories from multiple perspectives and dimensions, to offer you the real voices of entrepreneurs and to witness the power of technological innovations.
China’s AI chip industry is still in its infancy, but with the development of AI technology there is a large potential market for AI chips capable of meeting the demands of AI applications. The sixteenth episode of Discover the Rising Stars interviews Shang Huibin, Vice President of TensorChip, which won the Emerging Talent Award at the China’s 5G Industry Innovation Entrepreneurship Contest’s national finale, about TensorChip’s innovations in AI chip technology.
Tapping the Potential of China’s AI Chip Industry
TensorChip was founded in 2019 on the principle of “Algorithm — Chip — Memory-computing Coordination” to provide enterprises with “AI + Business Algorithm” chips and solutions.
In 2020 the global AI chip market was valued at over US$ 10.1 billion, and according to projections by Jazzyear, China’s domestic AI chip market value could reach CNY 55.7 billion by 2023. With the arrival of super data centers, the popularization of AI applications, and the development of cloud — edge — end user integration, China’s AI chip enterprises have strong market potential.
“AI technology is currently trending, but computing power depends on chips, and domestic chip utilization has much room for improvement,” Shang Huibin said.
“Right now there are three different development models for the domestic chip industry. The first is to make GPU products like NVIDA (the inventor of the graphics processing unit or GPU) as a domestic alternative. But because the market is dominated by monopolies and because it’s hard to keep up with NVIDA in terms of functionality and processing power, this route is difficult. The second is support or accelerate common operators extracted from mainstream algorithms. But chips made for old operators aren’t compatible with the new operators. The third is our approach, that is, we took into consideration the disadvantages of the previous two models and developed the reconfigurable memory-computing AI chip.” Shang Huibin told us.
As a core technological innovation in China’s chip industry, TensorChip’s memory-computing technology and reconfigurable technology help overcome GPU industry barriers and the homogenous competition within big data AI.
China’s Leading Digital Memory-Computing and Reconfigurable Technology Chips
Using advanced reconfigurable memory-computing chips and technology, TensorChip’s chips have all the power and GPU flexibility of ASIC chips, while also having a higher price-performance ratio. With reconfigurable technology, algorithms can be upgraded while keeping costs down; while memory-computing architecture increases computing power and decreases latency, making it perfect for 5G and other low-latency AI computing tasks. Reconfigurable memory-computing processor (RMUs) bring together the flexibility of reconfigurable architecture with the power of memory-computing technology.
TensorChip Currently Has Three Core Products
The first is CloudCard, an AI Inference Calculation Card, which uses advanced memory-computing architecture with deep optimization of the memory wall and compile wall for stronger and more efficient large-model support. Its RMU chip uses hot hard compression, and the design of its high resolution processing components gives it an energy efficiency ratio of 15.2TOPS / W, or 10–15 times that of a traditional GPU, with flexible operator variable capability and high compatibility with deep-learning deployment environments.
The second is EdgeChip, which offers cost effective and low power-consumption AI inference, providing powerful computing support for all kinds of edge computing. It supports flexible algorithm deployment for edge computing and custom operators from clients, providing innovation for multi-mode and multi-scene edge computing.
The third is an AI Acceleration SDK, which provides one-stop deployment for hardware engineers with 3–500 times typical AI performance and friendly compatibility for open source ecologies. It supports ARM / RISC-V / X86 platforms and smooth migrations to clients’ own algorithms. TensorChip’s products are widely applicable in cloud computing, self-driving vehicles, smart security and other fields, and provide integrated chip computing services to help customers integrate their own algorithms and platforms for efficient large scale production.
TensorChip’s products have numerous advantages. For example, using the AI accelerator pack (tinyAI SDK) to deploy RISC-V core AIoT applications has been found to be 11.01 times faster than other AI algorithms for Nuclei UX 600 AI applications. These clear advantages led Nuclei to choose tinyAI as first SDK to support UX600 graphics AI computing and acceleration. The tinyAI SDK incorporates AI model pruning and customization, and is optimized for RISC-V core storage and architecture. TinyAI not only makes AI computing easier on RISC-V on-device platforms, it also greatly improves AI algorithm performance, and helps make on-device AI a reality.
Expert Technology Team Tapping the Boundless Potential of AI
TensorChip’s core team come from AMD, Renesas Semiconductor, MediaTek, Yangtze Memory Technology Corp., and other international leading companies, and have years of experience in the chip and hardware acceleration industries.
Co-founder Dr. Chen Wei has a B.A. and M.A. from Tsinghua University and a Ph.D. from the Chinese Academy of Sciences, and is an expert in memory-computing AI chips with over 50 patents in the field, while Shang Huibin worked for 12 years in the artificial intelligence and chip industries. TensorChip’s core team not only has academic training, but rich industry experience.
When asked about the challenges TensorChip is currently facing, Shang said that “AI computing takes place mostly in the virtual world of the cloud, but there are many application scenarios for AI in everyday life, as well as in edge environments where there are privacy concerns; and these kinds of needs can’t be met with traditional products.” But these difficulties also put Shang in mind of AI’s limitless potential, the privacy and security of intelligent user terminals, and the huge storage and data analysis capabilities of the cloud. He looks forward to the evolution of intelligent machines and the internet of things.
Currently TensorChip is engaged in AI cloud computing product prototyping and testing, and is working with its partners on internet, industrial intelligence, autonomous vehicle related projects. The success of TensorChip’s RMU is a sign of the rapid development of China’s chip industry, which is driven by high technology and patents. By finding practical development scenarios for advanced technology, China’s chip industry will continue gaining in strength, which is not only good for the industry, but for China’s efforts to become a high technology leader.
Deep Ties with Alibaba Cloud, Competition Success Allows a Better Understanding of Future Needs
TensorChip’s advanced AI chip technology has brought the company remarkable success in industry competitions. In 2020 at the 3rd China IC Innovation and Entrepreneurship Competition the company won second prize in the Angel Category, and was the only chip company among the four finalists at the 2020 Audi Digital Sales Innovation Contest. In 2021 TensorChip won the Emerging Talent Award at the China’s 5G Industry Innovation Entrepreneurship Contest’s national finale hosted by Alibaba.
Asked about the relationship with Alibaba, Shang said, “Last year we partnered with T-Head to work on artificial intelligence, and this year we participated in intensive accelerators in Shenzhen and Hangzhou sponsored by Alibaba. In addition the 5G industry competition has helped investors understand our projects and has helped us understand future customer needs. During the competition the committee provided very helpful advice for companies in the startup stage. Alibaba Cloud’s Innovation and Entrepreneurship Center has also been very helpful. As TensorChip grows, we hope we can continue working with Alibaba.”
In 2021, TensorChip has attracted millions of investment in its AI computing reconfigurable architecture chips, and Shang Huibin says that the company will continue researching core technologies, and will follow Chinese industry trends in using technology to create social value.