Alibaba Bets on RISC-V and Custom AI Chips in an Attempt to Differentiate and Dictate Its Future

In its attempt to differentiate from the fierce competition of the public cloud market and reduce its reliance on Western supply, 2019 is shaping up to be a watershed year for Alibaba. In July 2019, Alibaba unveiled its first Reduced Instruction Set Computer Five (RISC-V) Computing Processing Unit (CPU), Xuantie 910, at the Alibaba Cloud Summit in Shanghai, China. This was immediately followed by the launch of Hanguang 800, an Application-Specific Integrated Circuit (ASIC) for cloud Artificial Intelligence (AI) inference workload in September 2019 at the Apsara Conference 2019 in Hangzhou, China.

Xuantie 910 is a 16-core, 2.5 GHz CPU based on a 12 nm process node and boosts maximum single core performance up to 7.1 CoreMark/MHz. Alibaba introduces more than 50 instructions to Xuantie’s RISC-V instruction set to enhance the processor's arithmetic operations, memory access, and multicore capabilities. The technical specification for Hanguang, however, is thin. Alibaba claims that its computational efficiency is at 500 images per second per watt when benchmarked using ResNet-50.

Before being able to work on in-house chipset design, Alibaba acquired C-Sky Microsystems in April 2018 and partnered it up with its Academy for Discovery, Adventure, Momentum, and Outlook (DAMO Academy), Alibaba’s in-house Research and Development (R&D) subsidiary. As the first chipset company that received significant funding from Alibaba prior to the acquisition, C-Sky Microsystems has launched a wide range of embedded cores, System-on-Chip (SoC) platforms, software tools, and middleware, and is well entrenched in the RISC-V open-source ecosystem, as discussed in the ABI Insight Watch Out for “Made in China” Open Source Artificial Intelligence Chipsets (IN-5269). The company also partners with UltraSoC, a U.K.-based startup that offers SoC intelligent self-analytics capabilities. Following the acquisition, C-Sky Microsystems changed its name to Pingtouge. The series of announcements in 2019 is a demonstration of a fruitful acquisition.

The announcement has drawn comparison to Google’s Tensor Processing Unit (TPU) and Amazon Web Services’ (AWS’) Inferentia. While Alibaba’s Hanguang chipset strategy is in many ways similar to AWS’ Inferentia, a much better comparison would be Huawei’s Kunpeng and Ascend chipset. Huawei focuses on their full line-up of chipset and server hardware in this year’s Huawei Connect. The company introduced Atlas 300, 800, and 900, an AI-enabled Peripheral Component Interconnect Express (PCIe) card, server, and system respectively based on its Kunpeng 916 and 920 and Ascend 310 and 910 chipset. Both launched last year: Kunpeng series is a series of Advanced RISC Machine (ARM)-based SoCs targeted at cloud servers, while the Ascend series is an ASIC for AI.

Alibaba is very clear about the purpose of these two chipsets. Xuantie 910 will be deployed in high performance end devices, including 5G telecommunication, AI, and autonomous driving. On the other hand, Hanguang 800 chip is used exclusively by Alibaba to power their own business operations, especially in product search and automatic translation, personalized recommendations and advertising.Eventually, the chips will be introduced to clients of Alibaba Cloud.

Given the current uncertain climate, this is a clear move by Alibaba and Huawei away from Western-based chipset companies, particularly x86 technology in the cloud environment. According to ABI Research’s Application Analysis Report Cloud AI Chipsets: Market Landscape and Vendor Positioning (AN-5032), the cloud AI chipset market is expected to grow from US$3.5 billion in 2018 to US$10 billion in 2024, with cloud AI training chipsets accounting for US$6.1 billion and cloud AI inference chipsets US$3.9 billion, respectively. While chipset vendors such as NVIDIA and Intel remain dominant in this market, public cloud AI vendors, such as Alibaba, Huawei, Google, and AWS, are expected to capture 15-18% market share of the overall cloud AI market by 2024.

Aside from Alibaba and Huawei, Baidu also launched Kunlun, its own cloud AI chipset back in 2018. While other major Chinese cloud vendors, such as Tencent, Kingsoft, China Telecom, and UCloud, have yet to launch their own proprietary hardware, the Chinese chipset market has witnessed a significant amount of momentum recently. Cambricon Technologies, founded in 2016, is a Chinese Intellectual Property (IP) core licensing vendor that also designs and sells cloud AI chipsets. In 2018, Cambricon Technologies launched MLU100, which was adopted by Original Equipment Manufacturers (OEMs), system integrators, and AI developers, such as Inspur, Lenovo, Sugon, DiDi, iFlytek, and Hikvision.

This development is in line with the aim of the Chinese government to boost its “Made in China” strategy for AI chipsets, given the constant pressure from the United States. By encouraging Chinese companies to develop custom AI chipsets and invest in the RISC-V chipset ecosystem, Chinese AI developers will have a stronger grip over their AI technology roadmaps. Open source ecosystems like RISC-V will also allow Chinese AI developers to access large communities of developers and more extensive supplier and partner ecosystems, while not being held hostage by any Western supplier. The interesting question is whether or not China can ride this momentum to allow these AI chipset architectures to become the de facto standard for the industry, particularly in the case of RISC-V, in the long run.