Meet the New “Made In China” Chipsets – AI Chipsets for Natural Language Processing

Most of the hype around AI in China is related to machine vision. SenseTime, a Beijing-based visual recognition startup, has just become the world’s most valuable startup, with its local competitors not far behind. AI chipset startups such as Cambricon Technologies and Horizon Robotics that focus on edge deep learning chipsets have raised a significant amount of venture capital.

Interestingly, what is under the radar is chipset advancement related to natural language processing. Long operating in a walled garden market, Chinese webscale giants have released their own voice control home speakers. AliGenie (Alibaba), Xiaodu Zaijia (Baidu), DingDong (JD.com), and Tingting (Tencent) all support internet search, social media integration, smart home controller, online shopping, and ride-hailing, similar to their Western counterparts. However, all of them do not feature on-device AI processing. The training and inference of NLP models are still being performed in the cloud.

In June 2018, Rokid launched its own AI chipset. Known as Rokid KAMINO 18, the AI chipset is developed in conjunction with Hangzhou NationalChip Science & Technology, a Chinese chipset company. Established in 2014, Rokid is a Hangzhou-based Chinese startup that produces smart speakers with voice-based AI assistants. Rokid’s devices act as voice control front-end for simple functions, such as hands-free calling, alarm clock setup, and music streaming.

While the new Chinese chipset may not bring a significant boost in performance or capabilities, it does address one critical disadvantage of Rokid, that is to say the lack of a cloud ecosystem. Up against international giants such as Amazon, Apple, and Google, as well as domestic webscale companies such as Alibaba, Baidu, and JD.com, Rokid does not possess strength in cloud computing and database. All Internet search and query functions, for example, require integration with Baidu or Alibaba via Application Programming Interfaces (APIs). Having local or edge inference being performed using AI chipset reduces the reliance on cloud technology, brings down the latency, and opens new avenues for Rokid to implement more capabilities at the edge.

Despite how Rokid positions itself as a worthy competitor against webscale giants, such as Apple and Google, the company offerings come up short. Key functions that are supported by its competitors, such as social media integration, smart home controller, online shopping, and ride-hailing, are not found in Rokid. On the flip side, interaction with most of the aforementioned functions will be cloud based and will not leverage edge NLP chipsets. In addition, according to Rokid, the cost of KAMINO 18 is 30% lower than comparable products coming from ARM and MediaTek. This gives Rokid cost flexibility and, in the light of U.S. Commerce ban on ZTE, less reliance on Western technology suppliers.

The smart home market in China is fairly fragmented, with major domestic brands such as Haier, Midea, Gree Electric, and Xiaomi each having their own ecosystems. In recent years, NLP has also become somewhat of a novelty feature in Chinese consumer goods. Voice control features can be found in a wide array of products, ranging from typical smart home speakers and TV to air-conditioners, fridges, and kitchen hoods. Fierce competition within the consumer electronic industry means rapid innovation and quick time to market, providing ample opportunities for vendors to integrate NLP chipsets to enhance product features and capabilities.

Hangzhou NationalChip Science & Technology, the chipset manufacturer in this tie-up, has been in the chipset market for nearly two decades. The company’s main revenue comes from set-top box chipset. In the past two years, Hangzhou NationalChip Science & Technology has been focusing on the development of its GX8010 AI chipset, focusing on NLP. The company correctly identifies the similarity between a set-top box chipset and a NLP chipset. Both chipsets rely on signal demodulation, with the former focusing on audio and video encoding/decoding, while the latter focuses on audio signal processing. In addition, the algorithms used for audio and video encoding/decoding are similar to the deep learning neural network for NLP. After two years of research, Hangzhou NationalChip Science & Technology finally launched GX8010 at the 2018 International Consumer Electronics Show (CES). The chip is based on 40nm TSMC engineering, featuring Hangzhou NationalChip Science & Technology’s gxNPU (a Neural Processing Unit [NPU]), the ARM Cortex-A7 Central Processing Unit (CPU), and Cadence's Tensilica HiFi-4 Digital Signal Processor (DSP).

Aside from Rokid, there are several other companies that are launching NLP chipsets. Unisound is another startup that focuses on NLP, and it just launched its AI chipset, UniOne, in May 2018. The company claims that UniOne has been adopted by major Chinese consumer brands, such as Haier, Midea, and LeTV. Similar to its industrial peers, the chipset features CPU, DSP, and deep neural network framework. Recently, Baidu also launched its cloud-to-edge AI chipset, known as Kunlun. The chipset supports Baidu’s PaddlePaddle framework, which is the key foundation for Baidu’s Chinese virtual assistant platform DuerOS. Nonetheless, the commercial availability of Kunlun remains a question mark, and Baidu’s attempt to integrate Beijing-based high-end smart home speaker startup Raven Tech has not gone very well, proving that the smart home device business is still full of risks and challenges.

As mentioned earlier, the training and inference of NLP models are still being performed in the cloud and likely to remain so for the foreseeable future. However, there are also demands for voice control in other verticals. Outdoor vehicles, devices, and equipment lack cloud connectivity, which will impair their capabilities, response time, and user experience. NLP chipset provides the much necessary onboard inference capability that can offload the need for cloud-based inference. Given the demand for rapid innovation and feature integration, ABI Research believes that NLP chipset companies will emerge as key AI technology companies.