91³Ô¹ÏÍø

Generative AI has become pervasive, enhancing the capabilities of the devices we use for inspiration, ideas, and important decision-making in our daily lives. To enable AI, large volumes of data must be processed at a fast rate, driving the industry towards multi-die designs and the adoption of high-bandwidth, low-latency interfaces such as UCIe, LPDDR, UFS, and MIPI CSI. These interfaces must not only provide real-time connectivity but also meet system-level requirements for power, performance, and area efficiency. In this paper, we will explore the motivation behind the transition to multi-die designs and examine how die-to-die, memory, storage, and MIPI interfaces interoperate to support large-scale AI-enabled systems. Additionally, we will delve into various use cases of multi-die designs in edge AI applications.

Test coverage is an increasingly complex Design for Test (DFT) challenge. Until recently, DFT is implemented at late stage of the design process, with limited options to impact RTL decision-making. Evolving Safety applications have imposed higher requirements on both RTL design and test coverage as high as 99%. Per ISO26262 requirements, safety mechanisms, such as safety error code, need to be used during RTL design. The large AND/OR gate dominant structural logic is instantiated as part of safety error code. Such logic 0 or 1 dominant structure is not well interpreted by ATPG engines. The path sensitization requires huge amount of care bits within patterns, lowering ATPG efficiency, Test Point Insertion effectiveness and increasing pattern counts. Moreover, large number of inputs lead to random resistance (RR) logic and limit the test coverage of an IP subsystem to less than 95%. It hinders the overall SoC coverage per AEC-100 requirement of a safety chip. ATPG pattern generation efforts can be reduced exponentially by proper assessment of the underline conditions and addressing through RTL recoding. RTL recoding can significantly improve test efficiency and logic overhead by addressing the root cause through logic splitting and other strategies. This presentation includes case studies that aim to illustrate recoding methods for improving test coverage of RR logic from 94% to 99.92%, and other scenarios of interest within actual interface IPs examples.