Signaling its future growth trajectory, Marvell Technology has confirmed that the company is working on data infrastructure products enabled by technology platforms on 5, 3, and 2nm. While the company has teams across the world, India remains its largest R&D development center outside the US.
"Work related to this investment in 2nm has already commenced," said Sandeep Bharathi, Marvell's Chief Development Officer. "While we don't publicly disclose specific details on products and process technologies, I can confirm our investment in advanced process nodes, including 5/3/2nm, is underway. Our experts, including those in India, cover all domains – digital design, architecture, analog, mixed-signal layout, circuit design, verification, and validation.
"These teams are integral to our projects, whether they involve 5-, 3-, or 2-nanometer processes. India hosts our largest R&D development center outside the US, contributing significantly to our entire product portfolio."
Focus on AI-centric growth
As AI continues to make inroads into several industries, Marvell sees value addition in the increased need for data centers. With AI models relying on trillions of parameters, this would make sense.
"The data center is a crucial component of AI strategy, differing significantly from general-purpose data centers," Bharathi said. "This distinction is what we refer to as accelerated infrastructure. The acceleration goal here involves improving computing power, storage, and connectivity.
At the recently concluded VLSI Design conference in Kolkata, Bharathi highlighted the data center's pivotal role in AI strategy, setting it apart from general-purpose facilities and establishing it as the foundation of data infrastructure.
"Marvell's mission statement remains unchanged in the era of AI, as our focus has always been on infrastructure," Bharathi said. "Our goal is to move, store, process, and secure data more quickly, effectively, and reliably than anyone else. This was our mission statement as established by Matt Murphy when he became CEO, and AI merely accelerates these objectives."
Optimizing computational power
Then, there is the issue of computational power. Machine learning has existed since the 1950s, as seen in single-layer perceptron models. Its current feasibility is due to advances in semiconductors, enabling the integration of substantial CPU or GPU power along with considerable memory capacity in the models and facilitating gigabytes to terabytes of connectivity.
"In data centers, where connectivity power can constitute around 10% of energy usage, low latency, low bit error rate, and low power consumption are critical," Bharathi added. "Marvell focuses on these aspects to control energy costs effectively. We provide custom compute silicon for specific workloads, especially relevant in AI applications where operations like vector math, matrix multiplication, and gradient computation require focused processing."
Focusing on data centers, Marvell offers products for various connectivity ranges using technologies like optical DSP and coherent modulation, with notable products like NOVA and Orion.
"Marvell's distinction also lies in our advanced CMOS geometries," Bharathi said. "We have a wide range of products in five-nanometer technology and are investing in three-nanometer technology, a field where few companies excel. This is a key differentiator for Marvell in the industry, given the high costs and complexities of development in these advanced geometries."
Plans in India
India plays a crucial role in Marvell's strategic plans. Over the last four to five years, the company has significantly expanded operations in India across various domains.
"We have comprehensive teams in India, including software and validation teams, which play a vital role in validating platforms right after silicon is returned," Bharathi said. "Our presence in India has also grown through acquisitions that enhance our design capabilities. We have established teams in Pune, Bangalore, and Hyderabad, each contributing unique specialties. Hyderabad is notable for its software development, while storage, software, and chip-building capabilities are prominent in Bangalore and Pune."
Marvell has expanded its portfolio through strategic acquisitions, including Cavium, Aquantia, Avera, Innovium, InPhi, Tanzanite, and SpicaWorks, enhancing its capabilities in critical areas such as processor, custom ASIC, network, and switch design. The contributions from its Indian operations remain pivotal to these enhanced capabilities.
"Almost every product segment of Marvell has representation in India," Bharathi said. "Our objective is to position India as an equally important development site for all our data infrastructure products, mirroring the significance of our other global sites."
