Design and Development of High Density Fan-Out Wafer Level Package (HD-FOWLP) for Deep Neural Network (DNN) Chiplet Accelerators using Advanced Interface Bus (AIB)

Emerging applications such as machine learning (ML) and artificial intelligence (AI) require more computing capabilities that ought to be distributed and have access to large memory and storage, while the systems need to be energy efficient and low-cost. The increase in cost of advanced nodes and the difficulties of shrinking analog circuits such as input and output (I/O) to address the computation and communication needs of ML/AI applications have created the opportunity to bring into the mainstream chiplet-based systems. The chiplet based systems enable modularity, scalability and technology partitioning providing a cost and energy efficient solution. The chiplet integration has been enabled by the development of a raft of advanced packaging technologies such as silicon interposer, EMIB, COWoS, high density fan-out wafer level packaging (HDFOWLP) to name a few. In this work the design, development and electrical characterization of a four-chiplet system integrated using in 2.5D HD-FOWLP platform is discussed. The chiplet accelerators are fabricated in 22 nm CMOS technology, while the package uses a five metal layer HD-FOWLP with dielectric polymer and 2 um width and space as minimum design rules. The Advanced Bus Interface (AIB) die-to-die PHY-level standard is used to interconnect the four chiplets in a ring topology. The AIB bus requires 192 lines between each two chiplets, and a total of 768 2umx2um lines are routed on the top three layers of the HDFOLWP. The bottom two metal layers of the package are used to distribute the ground and power necessary for all four chiplets. Each chiplet requires seven distinct voltage islands that are separately routed on the bottom metal layer.