Since Vice President and COO Qi Lu announced his departure from Baidu a month ago, questions have arisen regarding the Chinese search engine giant’s commitment to its oft-repeated “All in AI” pledge. At the Baidu Create 2018 AI developer conference which kicked off today in Beijing, the company announced a series of AI-based innovations and product releases that seem designed to regain public trust.
At last year’s inaugural Baidu Create, Founder and CEO Robin Li rolled out open-source AI platforms Apollo, billed as “the Android of Auto”; and DuerOS, the company’s Alexa or Siri-style virtual assistant, announcements in line with China’s determination to develop autonomous self-driving and virtual assistant technologies.
Synced is at the China National Convention Center in Beijing to bring you Baidu Create 2018 news and updates.
China’s first edge-to-cloud chip with 260 TOPS
The Internet giants’ race to create custom AI chips has so far produced Google’s powerful TPUs and Microsoft’s FPGAs. Two months ago, China’s Alibaba announced a development plan for its new AI inferencing neural network chip Ali-NPU. Today, it was Baidu’s turn.
The biggest wow moment on Baidu Create’s opening day came when the company announced China’s first-ever edge-to-cloud chip Kunlun (昆仑), which includes the training chip “818-300” and the inference chip “818-100” and can be applied to both cloud and edge scenarios, including data centers, public clouds, and autonomous vehicles.
In his talk Li boasted that Kunlun is “by far the most powerful AI chip in China’s industry.” Composed of thousands of small cores with 14nm Samsung engineering and 512 GB/second memory bandwidth, Kunlun delivers performance of 260 TOPS while consuming 100 Watts of power.
The Kunlun chip’s evolution can be traced back to 2011, when Baidu began developing an FPGA-based AI accelerator for deep learning and deployed GPUs in data centers. Kunlun can empower computation in multiple AI applications, including voice recognition, search ranking, natural language processing, autonomous driving, and large-scale recommendations.
Baidu says it will iterate upon this chip, developing it progressively to enable the expansion of an open AI ecosystem. There was no indication as to when or whether Kunlun will be made available to the public.
China’s first L4 autonomous bus
Earlier this year at the Baidu World Conference in Las Vegas, Baidu announced its autonomous driving solution Apollo would be incorporated into Chinese manufacturer King Long’s self-driving buses on designated public roads by the end of July 2018.
Today, Baidu kept its promise, announcing it has begun mass production of Apolong(阿波龙), China’s first L4 fully autonomous bus. Developed in partnership with King Long and powered by Apollo, Apolong will be put into commercial operation in Beijing, Shenzhen, Xiongan, Wuhan, and Pingtan of Fujian province. Baidu is also keeping an eye on overseas markets.
Baidu also announced that it has teamed up with SB Drive, the autonomous driving subsidiary of Japan’s SoftBank Group, to bring Apolong autonomous minibusses to Tokyo in early 2019.
So far over 100 Apolong buses have rolled off the production line in Xiamen.
Released in July 2017, Apollo is Baidu’s open-source platform aimed at democratizing autonomous driving. It provides developers access to a complete set of vehicle, hardware, software, and cloud data service solutions; as well as an API and codes for obstacle perception, route planning, vehicle control, and operating system.
In January Baidu said that over 165,000 lines of code had been created on Apollo, with an average of 146 “Commits” per week received from developers on Github. Today, the company announced the number has ramped up to more than 220,000 lines of code.
Also announced today was the upgraded Apollo 3.0, with improved support for autonomous driving in geo-fenced areas.
Apollo plans to work with Intel subsidiary Mobileye to integrate and commercially deploy Mobileye’s Responsibility Sensitive Safety (RSS) model, a formal model for safety assurance in Automated Vehicle (AV) decision making.
Baidu Brain 3.0
Introduced in September 2016, Baidu Brain is the baseline for Baidu AI services from image recognition to natural language understanding. Baidu Brain 3.0, announced today, expands the scope of its AI services and significantly improves its semantic understanding.
Haifeng Wang, Baidu Senior Vice President and Head of Baidu AI Group, explained: “The core [of Baidu Brain 3.0] is its multimodal deep semantic understanding. This means that for modal information such as images, voice, and video, machines can not only hear and see, but also understand the meaning behind it.”
Joining the Baidu Brain upgrade is DuerOS 3.0, a new version of the company’s platform for conversational AI. DuerOS supports home appliances like TVs and smart speakers and mobile devices like phones or watches. Developers can access open-source SDKs and APIs to build third-party voice conversational services. The DuerOS bot platform also provides many skills. Over 90 million devices have been embedded with DuerOS.
The DuerOS 3.0 system now can enable conversational capabilities on Bluetooth devices.
A trending Chinese Internet innovation is “Mini-programs,” which typically are less than 10 megabytes and offer app-like features on an application’s interface. Tencent’s message app WeChat is China’s largest mini-program host, accommodating one million and counting.
Baidu is testing the market by launching smart mini-programs in the Baidu app. Baidu smart mini-programs are fully supported by Baidu Brain 3.0, developers can call it through a few lines of code. Baidu will open source the smart mini-programs this December.
Baidu is on the right track, following its self-prescribed AI development steps. The company has begun to transform its AI research results into marketable products and services. The exciting Kunlun chip meanwhile sends the message that Chinese internet companies are taking semiconductors seriously. As the age of AI evolves, tech giants are broadening their business, from Internet to AI chips to smart products.
Journalist: Tony Peng | Editor: Michael Sarazen
merci pour les informations