According to Digitimes they are expected to make a killing in 2018 from their launch of various AI (artificial intelligence) chipsets, mostly ASICs and NPU (neural processing unit) chips, for a variety of applications in the second half of 2017.
TSMC will make the chips and its AI chip foundry services are expected to grow exponentially in 2018 along with volume shipments of the AI chips by the China makers, according to industry sources.
Huawei's HiSilicon chip design arm worked out the Kirin 970 as the new flagship SoCwith built-in AI computing capabilities and it was adopted in Huawei's Mate 10 and M10 Pro smartphone models launched in mid-October 2017. Official production of the Kirin 970 chips kicked off in mid-2017 using TSMC's 10nm FinFET process at a monthly capacity of 4,000 pieces of 12-inch wafers, placing Huawei among TSMC's top-5 customers.
Huawei is working on sprucing up AI capabilities on smartphones and wants to capture a 40 percent share of the China smartphone market. Huawei requires stable and sufficient supply of AI chips and even has to seek second supply sources other than TSMC, the sources said.
Cambricon Technologies released three new AI processor IPs in November 2017: the Cambricon-1H8 for lower consumption computer vision application, the higher-end Cambricon-1H16 for more general-purpose applications, and the Cambricon-1M autonomous driving applications.
While licensing AI processor IPs to end device vendors, Cambricon is selling chips to those in the cloud market. The company has newly debuted MLU100 AI chips to support inference application by datacenters and small- to medium-size servers, and MLU200 chips to support training applications at AI R&D centers of enterprises. These two AI chips will be manufactured using TSMC's 16nm process.
Horizon Robotics officially rolled out two Gauss-based AI processors, 1.0 Journey and 1.0 Sunrise, in December with the former for image processing and the latter supporting smart city applications with low power consumption. The company plans to introduce Bernoulli-based processor in 2018 and Bayes-based processor in 2019 with higher-performance AI chips.
Horizon has recently raised around US$100 million in series A+ financing led by Intel Capital to support its development of a prototype driverless car and driving technology innovations. The company aims to have its AI chips applied to more than 100 million IoT (Internet of Things) devices by 2020 before realizing the goal as a leading supplier of autonomous driving chipset solutions by 2025.
DeePhi Tech plans to debut two system chipsets in 2018, one for AI cloud services and the other for AI terminal devices applications, with the latter to adopt the firm's in-house-developed Aristotle architecture and manufactured using TSMC's 28nm process.