Fengda Sun
The emerging three-dimensional (3D) integration technology is expected to lead to an industry paradigm shift due to its tremendous benefits. Intense research activities are going on about technology, simulation, design, and product prototypes. This thesis work aims at fabricating through-silicon vias (TSVs) on diced processor chips, and later bonding them into a 3D-stacked chip. How to handle and process delicate processor chips with high alignment precision is a key issue. The TSV process to be developed also needs to adapt to this constraint. Four TSV processes have been studied. Among them, the ring-trench TSV process demonstrates the feasibility of fabricating TSVs with the prevailing dimensions, and the whole-through TSV process achieves the first dummy chip post-processed with TSVs in EPFL although the dimension is rather large to keep a reasonable aspect ratio (AR). Four self-alignment (SA) techniques have been investigated, among which the gravitational SA and the hydrophobic SA are found to be quite promising. Using gravitational SA, we come to the conclusion that cavities in silicon carrier wafer with a profile angle of 60° can align the chips with less than 20 µm inaccuracies. The alignment precision can be improved after adopting more advanced dicing tools instead of using the traditional dicing saws and larger cavity profile angle. Such inaccuracy will be sufficient to align the relatively large TSVs for general products such as 3D image sensors. By fabricating bottom TSVs in the carrier wafer, a 3D silicon interposer idea has been proposed to stack another chip, e.g. a processor chip, on the other side of the carrier wafer. But stacking microprocessor chips fabricated with TSVs will require higher alignment precision. A hydrophobic SA technique using the surface tension force generated by the water-to-air interfaces around the pads can greatly reduce the alignment inaccuracy to less than 1 µm. This low-cost and high throughput SA procedure is processed in air, fully-compatible with current fabrication technologies, and highly stable and repeatable. We present a theoretical meniscus model to predict SA results and to provide the design rules. This technique is quite promising for advanced 3D applications involving logic and heterogeneous stacking. As TSVs' dimensions in the chip-level 3D integration are constrained by the chip-level processes, such as bonding, the smallest TSVs might still be about 5 µm. Thus, the area occupied by the TSVs cannot be neglected. Fortunately, TSVs can withstand very high bandwidths, meaning that data can be serialized and transmitted using less numbers of TSVs. With 20 µm TSVs, the 2-Gb/s 8:1 serial link implemented saves 75% of the area of its 8-bit parallel counterpart. The quasi-serial link proposed can effectively balance the inter-layer bandwidth and the serial links' area consumption. The area model of the serial or quasi-serial links working under higher frequencies provides some guidelines to choose the proper serial link design, and it also predicts that when TSV diameter shrinks to 5 µm, it will be difficult to keep this area benefit if without some novel circuit design techniques. As the serial links can be implemented with less area, the bandwidth per unit area is increased. Two scenarios are studied, single-port memory access and multi-port memory access. The expanded inter-layer bandwidth by serialization does not improve the system performance because of the bus-bottleneck problem. In the latter scenario, the inter-layer ultra-wide bandwidth can be exploited as each memory bank can be accessed randomly through the NoC. Thus further widening the inter-layer bandwidth through serialization, the system performance will be improved.
EPFL2011
