Marvell Industry Analyst Day Showcase

Will Chu, SVP and GM, Custom Cloud Solutions BU, discusses the company's expansion into XPU attach solutions, where Marvell customizes all components within the XPU tray, including CXL-enabled memory for expansion and near-memory compute, security devices for AI infrastructure management, and high-performance NICs.

Radha Nagarajan, SVP and CTO, Optical and Cloud, outlines three networking architecture categories and how Marvell addresses all three segments through its Celestial acquisition for scale up, internal development for scale out, and an established business in scale across, including 1.6 terabit per second single wavelength links currently in development.

Xi Wang, SVP and GM for Connectivity, introduces the company's RELIANT software suite and Golden Cable program designed to remotely monitor and manage AI infrastructure connecting hundreds of thousands of GPUs and XPUs by collecting data from DSPs in cables and optical modules.

Matt Kim, AVP, Custom Cloud Solutions Business Unit, presents PIVR (Package Integrated Voltage Regulator) technology as a solution to power delivery challenges in the multi-kilowatt chip era where moving voltage regulators directly into the XPU package increases current density performance by up to 2x and reduces power transmission losses by as much as 85%.

Annie Liao, Product Management Director outlines Marvell's PCIe retimer product roadmap from generations 5.0 through 8.0, explaining how current electrical solutions will face increasing challenges at higher data rates of 128 and 256 Gbps per lane. She describes their strategic development of optical PCIe solutions, including partnerships for optical modules and cables as well as in-house co-packaged optics (CPO) technology.

Rishi Chugh, VP and GM, Data Center Switching at Marvell, joins Roy Chua, Principal and Founder of AvidThink, to explore how AI workloads are fundamentally changing data center networking by demanding flatter architectures and unprecedented levels of reliability. The discussion covers how AI training's massive power consumption and time requirements make network failures extremely costly, driving the industry to reconsider traditional Ethernet approaches and adopt more complex designs with distinct front-end and backend networks.