Broadcom has begun shipping what it calls the industry’s first 2 nm custom compute system on chip built on its 3.5D extreme dimension system in package platform, positioning the technology as a key building block for next generation xpue architectures. The 3.5D xdsip platform is described as a modular, multi dimensional stacked die approach that combines 2.5D techniques and 3D ic integration using face to face technology to increase integration density and performance.
The company frames 3.5D xdsip as foundational to future xpues designed for consumer artificial intelligence workloads, with the platform intended to deliver very high signal density, improved power efficiency and low latency. These characteristics are aimed at meeting the computational demands of gigawatt scale artificial intelligence clusters, where performance per watt and interconnect bandwidth are critical design constraints.
Broadcom states that the xdsip platform enables compute, memory and network input or output to scale independently within a compact package, which is intended to support high efficiency, low power computing at large scale. By separating and stacking these functional components in a 3.5D configuration, the approach is designed to allow system designers to tailor resources for specific artificial intelligence use cases while maintaining tight integration and minimizing physical footprint.
