Microsoft is building a network card for its own data centers. The card is primarily intended to improve the efficiency of clusters with GPU servers.

Microsoft likes to buy GPUs from market leader Nvidia, but when it comes to network cards, the Azure team prefers to take matters into their own hands. Nvidia has its ConnectX 7 series of network adapters and Microsoft is now said to be developing an alternative internally. Microsoft has not yet confirmed this itself, but it was passed on to The Information by an anonymous source. It could take more than a year for Microsoft’s development work to bear fruit.

Relieve CPU load

AI clusters in particular should benefit from their own network card. In this way, Microsoft hopes to be able to use its GPU servers more efficiently. Such servers work together like cloud supercomputers for complex tasks such as AI training. This collaboration is coordinated, also through shared memory. The network communication for this is essential, but is usually carried out via the CPU. This is slow and puts a strain on the valuable capacity of the GPU servers. Your own network card can only take care of the network utilization, so that it is optimized and the processor is relieved at the same time.

Since such a controller handles network traffic independently of the CPU, there are further advantages. The network card itself can incorrectly encrypt data, which in turn reduces the load on the CPU. All of these tasks are possible with the Nvidia ConnectX-7 and Microsoft probably wants to achieve the same with its own hardware.

Microsoft builds several data center chips in-house. Given the size of the company and the Azure cloud, this labor-intensive (and costly) work can pay off. Given Microsoft’s focus on AI, it makes sense that the company would want to take on certain tasks. The development of our own network card certainly fits into this picture.