Navin Bishnoi, Marvell – As billions of users simultaneously interact with AI platforms  like ChatGPT, the challenge extends beyond simply providing accurate responses. It  involves ensuring that these responses are delivered instantly, securely, and reliably. At  the heart of this seamless experience lies a crucial yet often overlooked component:  data centre networking. According to Navin Bishnoi from Marvell, specially designed  networking chips are now the backbone of extensive AI infrastructure. 

Marvell’s Contribution to the AI Ecosystem 

Marvell is a prominent global semiconductor company dedicated to the movement,  storage, processing, and security of data on a large scale. Its silicon solutions are widely  adopted by leading cloud providers, hyperscalers, and original equipment  manufacturers (OEMs). From cloud computing and AI workloads to carrier  infrastructure and enterprise networks, Marvell’s technologies are deeply integrated  into the data stack. 

The company boasts a diverse portfolio that encompasses compute, interconnect,  network switching, security, and storage, offered through merchant silicon, semi custom, and fully custom designs. This adaptability enables Marvell to cater to the  specific needs of modern AI-driven data centres. 

India’s Emerging Role in AI Infrastructure 

In India, Marvell is at the forefront of designing customised semiconductors, networking  switches, storage, and connectivity solutions. The company has established a strong

presence in India’s rapidly expanding AI infrastructure market, which is projected to  reach nearly $94 billion by 2028. 

Marvell collaborates closely with global hyperscalers and cloud providers, facilitating  everything from routine digital transactions to extensive AI model training. Its custom  silicon is embedded across compute systems, networking fabrics, and security layers,  making it a vital component of contemporary digital operations. 

Traditional vs AI Data Centres: A Shift in Networking 

While traditional data centres primarily focus on predictable data movement between  servers and storage, AI data centres operate on a vastly different scale. They  interconnect thousands—sometimes even hundreds of thousands—of AI accelerators  (XPUs) that constantly exchange large volumes of training data. 

In the “scale-up” scenario, hundreds of XPUs function collaboratively like a single  supercomputer. Here, ultra-low latency is crucial, as even a slight packet loss can  hinder AI training. In the “scale-out” scenario, high-speed connectivity, congestion  control, lossless Ethernet, and intelligent traffic distribution are essential to sustain  performance at scale. 

How Marvell Ensures Low Latency and High Bandwidth 

Marvell meets these challenges head-on with advanced switch architectures  specifically designed for AI and high-performance computing. Its Teralynx switches are  engineered to minimise latency during traffic surges, provide high bandwidth, and  support dense, energy-efficient network designs. 

By shortening the time data spends within the network, these switches contribute to  reduced overall AI job completion times, which is vital for large distributed training  tasks. 

Tackling Congestion and Traffic Spikes 

Network congestion poses a significant risk in AI data centres. Marvell proactively  addresses this by embedding intelligence directly into its switches. The system  continuously monitors traffic, detects early signs of congestion, and implements  corrective measures in real time. 

Shared memory pools offer additional buffering during sudden traffic surges, while  priority-based traffic management ensures that critical workloads receive the necessary  bandwidth. Fine-grained bandwidth control across ports and queues further guarantees  stable and predictable performance. 

Security in the Era of AI Infrastructure 

As AI systems proliferate across data centres, edge locations, and APIs, the risks 

associated with cybersecurity increase. Marvell integrates security directly into its  hardware, safeguarding AI models against threats such as data poisoning, theft, and  evasion. 

Its solutions encompass secure boot, robust authentication, encryption of data both in  transit and at rest, and confidential computing. The company is also preparing for future  security challenges posed by quantum computing, adopting a Zero Trust approach and  continuous threat monitoring. 

Power, Heat, and Engineering Challenges 

AI workloads exert tremendous pressure on power and cooling systems. Dense  networking and continuous processing generate significant heat, making energy  efficiency a paramount concern. 

Marvell tackles this challenge by optimising power consumption at every level—from  circuit design and architecture to software adjustments—helping data centres manage  heat while sustaining performance. 

In today’s AI-centric landscape, networking has evolved beyond mere support  infrastructure. It now serves as the foundation that enables intelligence at scale, and  Marvell is at the forefront, creating the silicon that keeps this ecosystem thriving.