Page 37 - ChipScale_May-June_2020-digital
P. 37

approximately 50% higher bandwidth and 100% additional capacity  Proven


        boasts a 3.2Gbps data transfer speed per pin, which is 33% faster
        than the previous-generation HBM2. Flashbolt has a density of
        16Gb/die, double the capacity of the previous generation. With these
        improvements, a single Samsung HBM2E package will offer a 410
        gigabytes-per- second (GBps) data bandwidth and 16GB of memory.
        In August 2019, SK Hynix announced an HBM2E DRAM product
        with the industry’s highest bandwidth. The new HBM2E boasts an

        compared to the previous HBM2. It supports an over 460 Gigabyte
        (GB)-per-second bandwidth based on the 3.6Gbps speed performance   Used in 5 Billion SiP Packages
        per pin with 1,024 data I/Os. By using through-silicon via (TSV)   in the past 5 years
        technology, a maximum of eight 16 gigabit chips are vertically
        stacked, forming a single, dense package of 16GB data capacity.
        Compared with traditional wire bond connections, DRAM chips that
        are stacked using TSV interconnects result in a shorter signal path
        and a high-speed performance with lower power consumption. While
        traditional structures package memory chips into a module form that
        can be connected to system boards, an HBM chip is packaged closely
        to logic chips, such as GPUs, leading to shorter distances between
        chips, which further accelerates data processing rates.

        3D packaging for AI
          Various 2.5D/3D packaging solutions, with or without TSV, are
        available today for packaging AI chips for inference and training.
        As shown in [1, Figure 1], TSV-based technologies include 2.5D
        Si interposer (e.g., chip-on-wafer-on-substrate [CoWoS]), 3D TSV
        stacking (e.g., for HBM stack), Foveros, 3D SoC, etc. TSV-less
        packaging technologies include embedded multi-die interconnect
        bridge (EMIB), integrated fan-out package on package (InFO-PoP),
        integrated fan-out on substrate (InFO-oS), fan-out chip-on-substrate
                                             ®
        (FOCoS), silicon wafer integrated fan-out SWIFT , integrated thin-
                                         ®
        film high-density organic package (i-THOP ), RDL interposer, etc.
        The positioning of various 3D packaging technologies in terms of I/O
        and package size is shown in [1, Figure 4].
          3D/2.5D TSV and heterogeneous integration technologies have
        emerged as the choice technology for AI, and particularly for deep
        learning applications, as they provide higher bandwidth, low latency,
        and low power consumption. For AI accelerators, it is important to
        keep logic and high-capacity memory as close as possible to provide
        low latency and lower power. When two chips or more are integrated   From water-soluble to no-clean processes,
        on an interposer, the distance between logic and memory is shortened,   Indium Corporation’s wide portfolio of
        which enables lower latency and lower power consumption. DRAM,
        based on a 3D TSV solution, offers an unequaled bandwidth   solders are PROVEN to solve various
        performance because of the ability of the TSV solution to connect   industry challenges.
        several layers of the device.
          3D IC and 2.5D solutions based on interposer are not stand-alone
        products—they need to be integrated onto a final package to be   Learn more:
        functional. The interposer acts as an intermediate layer between the   www.indium.com/SiP/CSR
        dies; it is the solution that has enabled die partitioning. Si interposers
        provide a high-density routing connection between logic and memory.   askus@indium.com
          Interposers can be made out of silicon, glass and organic laminate
        materials. Only silicon is used in commercial applications, but glass
        and laminate are under development. However, silicon interposers
        are the only solution ≤1µm for routing today. Silicon interposer is
        considered costly, so many alternative technologies from outsourced
        semiconductor assembly and test suppliers (OSATS) and integrated
        device manufacturers (IDMs) are in development, and some are
        already in production, e.g., EMIB from Intel, InFOoS from TSMC,
        FOCoS from ASE, etc. Some of the recent developments in the 3D
        stacking technologies are listed in the sections below.


                                                                                                             35
                                                               Chip Scale Review   May  •  June  •  2020   [ChipScaleReview.com]  35
   32   33   34   35   36   37   38   39   40   41   42