Marvell Unveils Breakthrough Co-Packaged Optics (CPO) Architecture for Custom AI Accelerators: Revolutionizing AI Server Performance
Marvell Technology, Inc. (NASDAQ: MRVL), a leader in data infrastructure semiconductor solutions, has announced a groundbreaking advancement in its custom AI accelerator (XPU) architecture with Co-Packaged Optics (CPO) technology. This innovation enables cloud hyperscalers and enterprises to scale up AI server capabilities dramatically, supporting hundreds of XPUs across multiple racks with higher bandwidth, longer reach, and optimal power efficiency. By integrating CPO into custom XPUs, Marvell is redefining the performance and scalability of AI infrastructure, addressing the growing demands of artificial intelligence workloads.
Revolutionizing AI Server Connectivity with CPO
Traditional AI servers rely on copper interconnects, limiting XPU-to-XPU connections to tens of devices within a single rack. However, Marvell’s new custom AI accelerator architecture with CPO overcomes these limitations by enabling hundreds of XPUs to be interconnected across multiple racks. This breakthrough is achieved through the integration of optical components directly into the XPU package, eliminating the need for electrical signals to travel through copper cables or printed circuit boards.
The result? Faster data transfer rates, 100X longer connection distances than traditional electrical cabling, and significant improvements in latency and power dissipation. These advancements allow cloud hyperscalers to build scale-up AI servers that deliver unparalleled compute density and performance while maintaining energy efficiency.
How Marvell’s CPO Architecture Works
Marvell’s innovative architecture combines XPU compute silicon, high-bandwidth memory (HBM), and other chiplets with its proprietary 3D Silicon Photonics (SiPho) Engines on the same substrate. This design leverages high-speed SerDes, die-to-die interfaces, and advanced packaging technologies to create a seamless, integrated solution.
At the heart of this architecture is the Marvell 3D SiPho Engine, first demonstrated at OFC 2024. Supporting 200Gbps electrical and optical interfaces, the engine integrates 32 channels of high-speed interfaces, modulators, photodetectors, and other critical components into a single device. Compared to previous generations, this engine delivers:
- 2x the bandwidth
- 2x the input/output bandwidth density
- 30% lower power per bit
These capabilities make it an ideal building block for incorporating CPO into next-generation custom XPUs, enabling hyperscalers to achieve higher interconnect bandwidths and longer reach without compromising latency or power consumption.
Proven Expertise in Silicon Photonics
Marvell’s leadership in silicon photonics spans over eight years, during which the company has shipped millions of devices, logging more than 10 billion field hours of operation. Its COLORZ® data center interconnect optical modules have been widely adopted by leading hyperscalers to meet growing bandwidth demands between data centers. Leveraging this expertise, Marvell has developed a cutting-edge CPO architecture tailored for the unique requirements of custom XPU connectivity.
“Silicon photonics is vital for scaling accelerated infrastructure connectivity to address increasing bandwidth demands, interconnect distances, power consumption, and total cost of ownership,” said Radha Nagarajan, senior vice president and chief technology officer of Optical Platforms at Marvell. “Since 2017, Marvell has pioneered the delivery of high-volume silicon photonics devices to top hyperscalers and leveraged this expertise to create a cutting-edge CPO architecture.”
Key Benefits of Marvell’s CPO Technology
- Higher Bandwidth Density: By integrating optics directly into XPUs, Marvell’s CPO architecture supports significantly higher bandwidth densities, enabling larger XPU clusters within a single AI server.
- Longer Reach Connections: Optical interconnects provide 100X longer reach than traditional copper cabling, allowing scale-up connectivity across multiple racks.
- Optimal Latency and Power Efficiency: The close coupling of optical components minimizes signal loss, enhances high-speed signal integrity, and reduces latency. Additionally, CPO eliminates the need for high-power electrical drivers, repeaters, and retimers, improving overall power efficiency.
- Enhanced Scalability: With CPO, hyperscalers can scale their AI servers from tens of XPUs within a rack to hundreds across multiple racks, unlocking new levels of compute performance.
Industry Recognition and Future Outlook
Industry analysts predict rapid adoption of CPO technology. According to Vlad Kozlov, founder and CEO of LightCounting, “We forecast that CPO will grow from less than 50,000 port shipments today to over 18 million CPO ports by 2029, with most deployments focused on server-internal connections.”
Marvell’s extensive experience in optical technology and custom XPUs positions the company as a key enabler of this transformation. “Integrating co-packaged optics into custom XPUs is the logical next step to scale performance with higher interconnect bandwidths and longer reach,” said Nick Kucharewski, senior vice president and general manager of the Network Switching Business Unit at Marvell.
Empowering Hyperscalers and Enterprises
The combination of Marvell’s custom AI accelerator architecture and CPO technology empowers hyperscalers to develop highly optimized AI servers capable of handling the most demanding workloads. By increasing XPU density and interconnect distances, Marvell’s solution addresses the scalability challenges faced by modern AI applications, from training massive neural networks to real-time inference tasks.
“Marvell’s custom AI accelerator with CPO architecture enables cloud hyperscalers to develop custom XPUs that will significantly increase the density and performance of their AI servers,” said Will Chu, senior vice president and general manager of the Custom, Compute and Storage Group at Marvell. “This innovation takes custom accelerated infrastructure to the next level of scale and optimization.”
Availability and Adoption
Marvell’s CPO-enabled custom XPU architecture is now available for customers’ next-generation designs. Multiple customers are already evaluating the technology for integration into their future solutions, highlighting its potential to become a cornerstone of AI infrastructure.
