Inspur continued its artificial intelligence (AI) and open data center push while contributing products and specifications to the Open Compute Project (OCP) at its regional summit in Amsterdam.

The company likes to brag that it’s the only server vendor that has joined all major open computing organizations including OCP, China’s Open Data Center Committee, and Open19.

One of its latest contributions is what Inspur calls the “world’s first” Open Accelerator Infrastructure (OAI) compliant and liquid cooling rack-scale AI computing product for deep neural network applications launched jointly with Baidu.

“This is the most exciting news,” said Alan Chang, deputy general manager of server product line at Inspur. The vendor expects to start shipping the new X-MAN 4.0 in early 2020. It addresses several challenges customers experience around implementing AI in their data centers such as hardware complexity, he added.

Baidu, Facebook, and Microsoft led the OCP’s OAI specification. It unifies the technical specifications of the accelerator module and simplifies the design complexity of the AI accelerator system, which in turn shortens time to market for the hardware system.

The X-MAN 4.0 is the first product built to this specification. It’s the fourth generation of Baidu’s X-MAN series for full-rack AI computing products. It includes eight AI accelerators in a single box, and it scales to 32 AI accelerators per rack. The boxes are interconnected through QSFP-DD cables to minimize latency across nodes.

The product also introduces OCP Accelerator Module (OAM) compliance accelerators, and the accelerator resources can be specified in a software-defined manner. All of this enables more flexible, multi-vendor support for various AI applications, Chang said.

The common form factor enables chipmakers like Intel and Nvidia to bring new silicon to market more quickly, Chang said. “The second benefit, for an [original design manufacturer] like us, we can consolidate our resources on focusing on how to optimize this design rather than saying I have to design a chassis for Nvidia, and another chassis for Intel,” he added. “And the end user benefit is: I want to do machine learning, or image recognition. Let me choose a couple options for silicon, put them on the same board, run a couple frameworks, and then design what works best for me.”

In addition to the product with Baidu, Inspur also announced the OpenRMC Version 1.0, North Bound Management Application Programming Interface (API) specification.

OCP first announced the Inspur-led OpenRMC (stands for “rack management module”) project at its Global Summit in March. It’s a subproject of the OCP Hardware Management Project, and it targets rack management controller specifications, requirements, and reference implementations.

The new OpenRMC NB API 1.0 is a four-socket, high-density server and rack-scale system designed for small- and medium-sized data centers. “One of the benefits of four socket,” as opposed to a two-socket system, “is the consolidation of the power supply,” Chang said. “We had one customer, in particular, saying they save from 60 watts to 80 watts.” This cost- and energy-savings helped the customer install one more server inside the rack, he added. “How good is that, and how much more computing can you get out of that by saving all of this wasted energy?”

Inspur plans to release OpenRMC versions 2.0 in March 2020, and says it will include a more complete rack reference system with a detailed WebUI.

Additionally, at the OCP Regional Summit, Inspur announced its latest OCP-accepted platform, NF8380M5. It’s a 3U, four-socket server based on the Project Olympus four-socket server, and it’s powered by Intel Xeon scalable processors and Intel chipsets.

The platform is designed for multi-scene AI applications and can drive GPU acceleration training workloads when serving as compute head node. It supports all non-volatile memory express solid state drive configurations for real-time decision database applications.

“This year, in 2019, we made a commitment to have three new contributions to the OCP market place, so people can buy them today and be fully compliant with the OCP standard,” Chang said. Inspur ended up beating this goal, and the new platform brings Inspur’s OCP-accepted product total to four.

And finally, Inspur contributed the I-Flex server design standard to the open source community. It’s based on the T-Flex 1.0, and it enables modularized deployment of servers based on different application scenarios to support server upgrades without changing the server architecture.