On November 6 in Beijing, China’s rising semiconductor company Cambricon released the Cambrian-1H8 for low power consumption computer vision application, the higher-end Cambrian-1H16 for more general purpose application, the Cambrian-1M for autonomous driving applications with yet-to-be-disclosed release date, and an AI system software named Cambrian NeuWare.
Company CEO Tianshi Chen also announced Cambricon’s plan to launch the MLU machine learning series processor for cloud-based applications in 2018.
So far, industry giants Intel and NVIDIA enjoy a strong lead in the AI chip race. A successful series of GPU products has pushed NIVIDA’s market value to over US$100 billion. Insiders like Will Yuan, Director of Strategic Business Development in Shanghai for British microprocessor leader Arm, believe the prohibitively high R&D and manufacturing costs of AI chips make this a marketplace only accessible to big semiconductor companies.
The high entry barrier has not however deterred audacious investing in chips. San Francisco startup Cerebras raised over US$112 million in three rounds of financing this year, while Wave Computing and Graphcore each raised US$60 million from investors including DeepMind’s Demis Hassabis, OpenAI, and Uber’s Zoubin Ghahramani.
Riding the global wave, China’s chip unicorn Cambricon attracted an impressive US$100 million in its Series A round this August, with much of the money coming from strategic investors like Alibaba and Lenovo. The current market value of Cambricon stands at US$1 billion.
In current computer architecture, CPUs are responsible for the general task and logic-controlled computing, while GPUs handle data-intensive, graphics-based vector tasks. However for AI processing which involves matrix multiplication and addition, both CPUs and GPUs fall short in instruction cycles. Cambricon’s specially designed NPUs (neural processing units) tackle this problem.
In 2016 Cambricon launched its AI processor “1A”, designed especially for mobile devices, with hugely augmented performance per watt. By the end of 2016, Cambricon had booked CNY 100 million worth of orders. Huawei’s newly-launched Kirin 970 chip is equipped with the 1A processor (NPU). The Kirin 970 chip installed on Huawei Mate10 has a HiAI mobile computing architecture comprised of four parts: CPU, GPU, ISP/DSP and a NPU.
Company CEO Tianshi Chen says Cambricon’s NPU “is at least two orders of magnitude” better than CPUs or GPUs for image and voice recognition.
In order to succeed in an arena with well-established competitors, Cambricon is tapping into the niche market of deep learning processors for specialized applications. These ASIC (application-specific integrated circuit) microchips can achieve exceptional performance in areas not yet covered by the industry giant’s generic processors.
However, custom ASIC chips can only run fixed algorithms, they cannot adapt to more general applications. Chen says he does not want Cambricon to be pigeonholed as an ASIC chip company: “Neural network processors is where we began, and we will be launching more general-purpose chips in the future.”
“We hope that Cambricon will soon occupy 30% of China’s IP market and embed one billion device worldwide with our chips. We are working side-by-side with and are on the same page with global manufacturers on this,” says Chen.
Journalist: Meghan Han, Duotian Yu (China), Zenan Li (China) | Editor: Michael Sarazen
Pingback: Synced | Unveiling China’s Mysterious AI Lead: Synced Machine Intelligence Award 2017
Pingback: AiYo Damo Namo: It’s On – Spider21's Weblog