Marvell Unveils CPO Architecture for AI Accelerators

At CES 2025, Marvell, a leading player in networking and optical communications, unveiled a significant breakthrough in its customized AI accelerator architecture. By integrating Co-Packaged Optics (CPO) technology, Marvell has significantly boosted server performance, enabling AI servers to scale more effectively. This new architecture allows AI servers to expand from dozens of XPUs within a single rack to hundreds of XPUs spanning multiple racks.

Marvell's CPO technology utilizes advanced multi-generation silicon photonics (SiPho), which has been in operation for over eight years and has accumulated over 10 billion hours of runtime. The integration of CPO promises data transfer speeds that are 100 times faster than traditional cables, positioning it as a transformative force in the AI industry. Industry giants such as TSMC, ASE Group, Intel, Broadcom, and Marvell are actively developing CPO technologies.

In a strategic move, Marvell has signed a five-year partnership with Amazon AWS, a global leader in cloud services, to supply customized AI chips. The integration of CPO technology into Marvell's AI chips is expected to drive a new wave of AI server deployments at Amazon AWS, accelerating their AI infrastructure development.

As the second-largest chipmaker in AI revenue worldwide, following Nvidia, Marvell is a key player in the customized AI chip market. The company is already poised to make 2025 the "Year of CPO," with its new AI chip architecture officially incorporating CPO technology. This innovation promises to revolutionize AI applications by delivering higher bandwidth and greater data transfer speeds, enabling more efficient, scalable AI systems.

Marvell's latest AI chip architecture combines XPU chips, high-bandwidth memory (HBM), and Marvell's 3D SiPho engine. By integrating optical technologies, Marvell can achieve 100 times faster data transfer and distance between XPUs than traditional cable-based connections. This advancement enhances the capabilities of AI servers, enabling the scaling of AI infrastructures across multiple racks, facilitating the development of custom XPUs with higher frequency and bandwidth density.

With Marvell's ongoing collaboration with Amazon AWS, the integration of CPO technology is set to accelerate the next generation of AI server infrastructure. As the first major cloud service provider to adopt these advanced chips, Amazon AWS is poised to lead the charge in this next phase of AI innovation.