Microsoft Challenges NVIDIA with AI-Powered Network Card

In recent years, NVIDIA has made significant inroads into the server market, especially with the augmentation of Artificial Intelligence (AI), achieving commendable success. Beyond its core products of CPUs and GPUs, NVIDIA’s acquisition of Mellanox and the subsequent integration of its networking product line have become indispensable to a myriad of data centers.

According to a report by The Information, Microsoft is developing its proprietary network solution, a custom network interface card (NIC) akin to NVIDIA’s ConnectX-7. This smart NIC, derived from the NVIDIA Quantum-2 InfiniBand architecture, is capable of delivering a throughput of 400Gb/s and supports NVIDIA’s network compute acceleration engine for additional speed, offering the high scalability and feature-rich technology required by supercomputers, artificial intelligence, and hyperscale cloud data centers.

The development of Microsoft‘s custom NIC is led by Pradeep Sindhu, co-founder of Juniper Networks, and is part of Microsoft’s broader strategy to tailor high-speed network hardware specifically for AI workloads. This initiative aims to expedite AI model training, reduce associated costs, and diminish reliance on NVIDIA’s networking products. Insiders have hinted that the development of the custom NIC could take just over a year, with the final design potentially offering up to 800Gb/s throughput.

In the long term, possessing custom network technology specifically engineered for AI could provide Microsoft with a competitive edge and greater control over optimizing and updating data center hardware. As AI permeates more business and consumer applications, Microsoft is making a strategic investment in next-generation infrastructure to support an AI-centric future.