Microsoft Corp.’s engineers are working on a custom network card for its data centers, The Information reported today.

A source told the publication that the development effort could take more than a year. The initiative is said to be led by Pradeep Sindhu, a co-founder and former chief executive of network equipment maker Juniper Networks Inc. That company is currently in the process of being acquired by Hewlett Packard Enterprise Co. as part of a $14 billion deal.

The servers in a data center are linked together by a network that allows them to share information with one another. Traffic doesn’t travel directly from a server to the data center network to which it’s attached, but rather goes through an intermediary known as a network card or adapter. It’s a specialized chip that helps route each packet to its destination and, often, performs related tasks such as optimizing connection speeds.

Microsoft reportedly plans to use its custom network card in its artificial intelligence infrastructure. In particular, the focus is on managing the network traffic of servers equipped with Nvidia Corp. graphics cards. Microsoft reportedly hopes that the custom card will help speed up AI workloads and reduce its hardware procurement costs in the process.

The Information’s source said the card is similar to Nvidia’s ConnectX-7 network adapter series. The latter product can process up to 400 gigabits of traffic per second. It supports both Ethernet and InfiniBand, the two communications standards on which most data center networks are based.

To coordinate their work, the servers that make up an AI cluster require the ability to share the data stored in their respective memory pools. Data sharing requests must typically go through a server’s central processing unit. Nvidia’s ConnectX-7 adapter includes a technology called RDMA that bypasses the CPU, which significantly speeds up data retrievals.

The device also features several other performance optimization features. Notably, it can perform cybersecurity tasks such as encrypting data traffic that would otherwise be carried out by a server’s CPU, which leaves more CPU capacity available for applications. ConnectX-7 also offloads some of the computations involved in detecting data transmission errors.

Microsoft may seek to match those features with its custom network card. Sindhu, the executive who is said to be leading the adapter’s development, joined the company in 2023 after it acquired a startup called Fungible Inc. he had founded. Fungible developed a chip that, similarly to Nvidia’s ConnectX-7, can offload certain cybersecurity tasks and related computations from a server’s CPU to boost application performance.

Microsoft’s reported network card would join the lengthy list of data center components it develops internally. The company recently detailed a custom AI accelerator, the Maia 100, along with an accompanying server rack and liquid cooling system it has likewise designed in-house. Microsoft plans to deploy the accelerator in its data centers alongside a CPU called the Cobalt 100 that its engineers created based on Arm Holdings plc designs. 

Photo: Microsoft