ChannelLife Australia - Industry insider news for technology resellers
Story image
Alibaba unveils new in-house server chips and releases source for IoT processors
Wed, 20th Oct 2021
FYI, this story is more than a year old

Alibaba Cloud has unveiled a new in-house processor designed for use in its data centers.

The new server chip, Yitian 710, has been custom-built by Alibaba Group's chip development business, T-Head. Alibaba Cloud has also announced the development of its proprietary servers powered by these chips, called Panjiu.

Alibaba Cloud Intelligence president and head of Alibaba DAMO Academy during the Apsara Conference, Alibaba's annual technology flagship event, Jeff Zhang, says customising their own server chips is consistent with ongoing efforts to boost computing capabilities.

 "We plan to use the chips to support current and future businesses across the Alibaba Group ecosystem," says Zhang. "We'll also offer our clients next-generation computing services powered by the new chip-powered servers in the near future.

"Together with our global partners including Intel, Nvidia, AMD and Arm, we'll continue to innovate our compute infrastructure and offer diverse computing services for our global customers."

Built on 5nm process technology, Yitian 710 is powered by 128 Arm cores with 3.2GHz top clock speed to deliver performance and energy efficiency. Each processor chip has 60 billion integrated transistors. Yitian 710 is the first server processor compatible with the latest Armv9 architecture and includes 8 DDR5 channels and 96-lane PCIe 5.0, providing high memory and I/O bandwidth.

According to Alibaba, Yitian 710 achieved a score of 440 in SPECint2017 (a standard benchmark to measure CPU integer processing power), surpassing the current state-of-the-art Arm server processor by 20% in performance 50% in energy efficiency.

Panjiu was developed for the next generation of cloud-native infrastructure. By separating computing from storage, the servers are optimised for both general-purpose and specialised AI computing, as well as high-performance storage. With a modular design approach for large-scale data center deployment, Alibaba says the servers are expected to deliver excellent economic value for a wide variety of cloud-native workloads, such as containerised applications and computed optimised workloads.
 
During the keynote, Alibaba also announced it would open the source code of the XuanTie IP core series, among other upcoming cores. The XuanTie series is Alibaba's custom-built processor based on RISC-V instruction-set architecture. Developers can now access the source code of the XuanTie series IP cores on GitHub and Open Chip Community to build prototype chips of their own, which can be customised for IoT applications such as networking, gateway and edge servers.

XuanTie related software stacks, which support multiple operating systems, including Linux, Android, RTOS and Alibaba's own AliOS, will also be opened. Alibaba says it will provide more services and support for development tools, SDKs and customised cores in the future.

"By opening up the IP cores of our in-house IoT processors as well as related software stacks and development tools, we aim to assist global developers in building their own RISC-V-based chips in a much more cost-effective way," says Zhang.

"We hope this move can encourage more innovation among the thriving RISC-V software community, and as a result, help people enjoy the benefits of a connected world in the digital era."

RISC-V International CEO, Calista Redmond, says Alibaba supports the RISC-V community through continuous contributions, technical leadership, and deep collaboration with RISC-V stakeholders.

"Alibaba leads by example and has inspired the global RISC-V community to increase innovation in chip development with benefits to the entire RISC-V ecosystem," Redmond adds.

Alibaba released its first AI inference chip, Hanguang 800, in 2019. The AI chip has been deployed in Alibaba Cloud's data centers to power businesses in areas such as search and live-streaming.