Page 35 - Chip Scale Review_January February_2022-digital
P. 35

Figure 3: Chiplet interface specification requirements for industry scale adoption.
        a monolithically-integrated alternative   can be individually, comprehensively,   communication latencies and improving
        (due to die size-based yield or reticle   and efficiently tested prior to assembly   system performance.
        size limitations, intellectual property   together with a chiplet interconnect and   In production systems and supply
        portability/suitability considerations,   packaging integration approach that   chains, any additions to cycle time
        development cost challenges to achieve   enables efficient and comprehensive post-  (processing and/or transit times) or
        all the required configuration variations,   assembly testing with an adequate degree   supply chain uncertainty will often
        etc.). Still, chiplets are not necessarily   of redundancy and repair can help mitigate   require additional inventory to be carried.
        going to be economically optimal in cases   this disadvantage. Likewise, not all chiplet   Purchasing external chiplets typically
        where a single “sweet spot” monolithic   solutions will necessarily require the   incurs margin stacking and additional
        alternative is feasible and attractive.  high bandwidth densities that advanced   inventory carrying costs. The use of a
          Continuing advances in packaging and   packaging can provide, so ideally, the   common set of chiplets in many chip
        interconnect technology will help reduce   chiplet interconnect selected can support   configurations can mitigate the impact
        tiling overheads over time. For example,   an option to use conventional low-cost 2D   of  these  considerations  by  enabling
        Intel’s Foveros Omni is projected to   “standard package trace” packaging when   inventory pooling. Likewise, consignment
        support bump pitches down to 25µm   that is economically optimal.     arrangements can help mitigate some of
        and Intel’s Foveros Direct is projected to   Thermal and mechanical constraints   the margin stacking and carrying cost
        support bump pitches of 10µm or less.   can also strongly influence system-level   impacts.
        Each N-fold improvement in bump pitch   partitioning. It is often impractical to   While 3D stacking can provide
        and wire density scaling for advanced   co-package multiple “hot” die together   substantial savings in board space,
                                        2
        packaging potentially enables an N    when doing so exceeds the thermal   interface density, and power efficiency,
        improvement in bandwidth density.   capabilities of the resulting system. This   chiplets built to be 3D stacked with
        Likewise, as process technology improves   limitation can be especially severe in   other chiplets have historically needed
        and connection distance shrinks, we can   those industrial or far-edge usages that   to be carefully co-designed (e.g., due to
        reduce the voltages chiplet interfaces   cannot rely on forced-air cooling. It is   thermal, power delivery, signal integrity,
        use to further improve energy efficiency.   likewise similarly impractical to co-  etc., considerations), hindering the ability
        Chiplet interconnects should therefore,   package die that each require a lot of   to reuse them in other contexts. However,
        ideally, use a “many wires, simple     external PCB connectivity because doing   as HBM memory has illustrated, we
        I/O” philosophy rather than a SerDes-like   so will increase package-level and PCB   can usefully specify a standard 2D/2.5D
        “few wires and complex I/O” approach   routing requirements and associated   interface to a 3D stack of chiplets that is
        to take better advantage of upcoming   costs. Instead, it is generally preferable   supplied as an integrated subassembly.
        improvements in connectivity technology.   to co-package functions that require   This consideration is a key reason
          A more subtle disadvantage of using   high-bandwidth connectivity between   why it is preferable to initially focus
        chiplets is that packaging, assembly, and   themselves to take full advantage of   multivendor standardization efforts on
        test costs and durations will generally   the previously noted 10x improvements   2D and 2.xD chiplet connectivity usages
        be higher for multi-die chips than   in connection density and energy   before attempting to standardize 3D
        monolithic chips. Using chiplets that   efficiency while incrementally reducing   interface details.

                                                                                                             33
                                                          Chip Scale Review   January  •  February  •  2022   [ChipScaleReview.com]  33
   30   31   32   33   34   35   36   37   38   39   40