After unveiling its proprietary 128-core datacenter CPU and Maia 100 GPU tailored for artificial intelligence workloads, Microsoft is now venturing into the development of its networking card to reduce dependency on Nvidia’s hardware and enhance the speed of its datacenters, as per a report by The Information. This strategic move aims to optimize Microsoft’s Azure infrastructure and broaden its technology portfolio. Interestingly, the company has indirectly acknowledged this initiative.
Approximately a year ago, Microsoft acquired Fungible, a company specializing in data processing units (DPUs) that rival AMD’s Pensando and Nvidia’s Mellanox divisions. This acquisition equips Microsoft with the essential networking technologies and intellectual property required to engineer datacenter-grade networking equipment suitable for bandwidth-intensive AI training workloads. Pradeep Sindhu, a co-founder of Juniper Networks and the visionary behind Fungible, now spearheads the development of Microsoft’s datacenter networking processors.
The forthcoming networking card is anticipated to enhance the performance and efficiency of Microsoft’s Azure servers, which presently utilize Intel CPUs and Nvidia GPUs, with plans to integrate Microsoft’s in-house CPUs and GPUs in the future. The Information asserts that this project holds significant importance for Microsoft, evident in Satya Nadella, the company’s CEO, personally appointing Sindhu to lead the endeavor.
A Microsoft spokesperson informed The Information, “As part of our systems approach to Azure infrastructure, we are focused on optimizing every layer of our stack. We routinely develop new technologies to meet the needs of our customers, including networking chips.”
Efficient high-performance networking equipment plays a pivotal role in datacenters, particularly when managing the substantial data volumes essential for AI training tasks by clients like OpenAI. By mitigating network congestion, this new server component could expedite AI model development, streamlining the process and enhancing cost-effectiveness.
Microsoft’s strategic move aligns with the prevailing industry trend towards customized silicon, mirrored by other cloud service providers such as Amazon Web Services (AWS) and Google, who are also delving into developing their own AI and general-purpose processors alongside datacenter networking equipment.
The introduction of Microsoft’s networking card has the potential to impact Nvidia’s sales of server networking equipment, which are forecasted to exceed $10 billion annually. Successful implementation of this networking card could notably boost the efficiency of Azure datacenters overall and enhance OpenAI’s model training specifically, while also trimming the time and expenses associated with AI development, according to the report.
Designing and manufacturing custom silicon entails a considerable time investment, indicating that tangible outcomes from this initiative may take several years to materialize. In the interim, Microsoft will persist in leveraging hardware from external vendors, though a transition is foreseeable in the upcoming years.
