As the market learning intensifies, several companies are trying to set themselves up with custom chips that can run inference operations in embedded devices instead of the cloud. Horizon Robotics is not only developing chips but also the software and cloud platform, with an eye toward prying into applications like security cameras and autonomous cars.
“The chip is the local brain that directly senses the surrounding environment, while the algorithm is the miner of the data,” said Kai Yu, founder and chief executive of Horizon Robotics, and the former head of Baidu’s artificial intelligence unit, called the Institute of Deep Learning, in an interview with Electronic Design.
“We want to empower end devices with A.I. capacity and make them smart without relying on the cloud alone,” Yu said, adding that the “chip and algorithms are used to perceive and filter big data, perform real-time processing and transmit valuable data to the cloud for further mining and modeling. Each component works together.”
The central component in Horizon’s SoCs is the brain processing unit, a custom block of circuitry that specializes in algorithms trained on vast libraries of images, hundreds of hours of video, or other data. The silicon slab can also be slipped into chips like field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs), Yu said.
The Beijing, China-based company is aiming to plant its chips into tens of millions of smart cameras over the next two years. Developing and deploying the technology so quickly led it to raise $100 million in venture capital last year from investors including Sequoia Capital, state-owned China Jianyin Investment, Harvest Fund Management and Intel Capital.
One of Horizon's chips, Sunrise, runs facial recognition or other inference algorithms with up to a trillion operations per second. The company recently released a security camera using it at the International Security Technology Show in Las Vegas. The camera can identify and follow 200 objects within each frame of video, recognizing the face of customers in a clothing store, for example, or plucking a criminal suspect from a crowded sidewalk.
With Sunrise, based on 40-nanometer technology, the camera runs at 30 frames per second while consuming 1.5 watts. With support for 50,000 different faces and 99.7 percent accuracy, the system can avoid the latency and bandwidth issues introduced by steaming data to servers in the cloud, where training and inference typically occur.
While Horizon has the upper hand over American rivals in China, the challenge is in keeping its hardware and software on the same page. While software engineers can change lines of code relatively fast, chip designers need several months and millions of dollars to prototype chips and get them back from the foundry. “It’s been a priority since the get-go,” Yu said.
He added that four-fifths of the company’s more than 300 employees have research and development backgrounds. Other founders of Horizon include Chang Huang, a founder of Baidu’s A.I. business unit, and Ming Yang, a founder of Facebook’s A.I. research team. Feng Zhou, a former principle chip architect for Huawei’s HiSilicon business, leads chip development.
In many ways, the company’s pincer attack on machine learning mirrors China’s national strategy. Not only has the country pledged $150 billion to close the technology gap between American and Chinese chip suppliers – and reduce its roughly $275 billion in annual chip imports – but it is also pushing to become the world leader in artificial intelligence by 2030.
The machine learning movement could level the playing field for China’s chips, particularly as the focus shifts to custom over commodity products. Horizon uses 40nm technology that entered production almost a decade ago, but it is jumping into embedded inference with companies like Qualcomm, which has started sampling 10nm chips that can be installed in networks of security cameras.
China’s ambitions include the deployment of 30 million autonomous cars within the next decade. Horizon is also trying to tap into the momentum behind that market. The company has partnered with Robert Bosch and Ford’s Chinese partner Chongqing Changan Automobile, among others, to put its automotive camera processor, called Journey, through its paces.
Horizon designed Journey to spot pedestrians, lane markings and other vehicles on the road and help driverless cars avert accidents. The chip may ultimately compete with systems from Santa Clara, California-based Nvidia and Intel’s Mobileye business. Bloomberg reported that the company is looking to have autonomous test cars on Chinese roads by 2019.
“China right now is a huge market for a lot of innovation happening in machine learning, from city management and security to autonomous driving,” said Yu, a member of the country’s strategic artificial intelligence advisory board. “China is a great playground for us to develop and mature these products for the rest of the world.”