![]() No effort was spent to optimize the circuits for performance. Consequently, density was not a primary objective of the physical design. The full-custom layout for the prototype was completed on a short time budget of ten days. As wafer alignment continues to improve, inter-tier via diameters will shrink, improving inter-tier connection density. Wafer assembly technology has advanced significantly since solder bumps on die surfaces were used to assemble multi-chip modules (MCM) vertically. Their inter-tier vias were etched (using inductively coupled plasma) through oxide and epoxy adhesive layers, lined in aluminum, but not tungsten-plugged due to temperature constraints. For comparison, the Rothko FPGA was fabricated with an older but similar SOI-based technology from Northeast- ern University, using wafer-scale transfers and assemblies (Kopin Technologies), featuring 6 μm diameter vias. have estimated the capacitance of these vias to be equivalent to that of a 8 μm - 20 μm wire (the exact value depends on the coupling to the surrounding ge- ometry). 5 μm wide, smaller than an SRAM cell in this technology, with a pitch of 5. (Typical 180 nm processes have six or more metal layers.) The inter-tier vias are 1. Each tier has three aluminum metal layers and tungsten- filled inter-tier vias. Tier 3 (top) is flipped and bonded to the back of Tier 2. Tier 2 (middle) is flipped and bonded to the face of Tier 1 (bottom). Figure 5 shows a cross-section of the three-tiers after they have been bonded. Using SOI wafers provides two benefits: i) the in- sulator acts as an etch stop when stripping away the substrate, and ii) since the substrate is completely re- moved, the via holes do not need to be passivated (insu- lated) thereby reducing the via pitch. Each tier is a fully-depleted SOI wafer bonded to other tiers via a low-temperature wafer-wafer oxide bond- ing. The 3D AFPGA is implemented in a three-tier, 180 nm silicon-on-insulator (SOI) process developed at the Mas- sachusetts Institute of Technology Lincoln Laboratory (MITLL). ![]() The scan chain itself was implemented with restor- ing logic as a conservative design choice, and consequently, accounted for a majority of the floorplan area. The logic block is configured by 96 bits, the 3D switch box is configured by 88 bits, and the connection box is configured by 20 bits. Our configuration logic consists of a four-phase non-overlapping clocked serial scan-chain with a single input and a single output for verification. In the future, we may see hybridization of these approaches. Other architectures exploit multi-layer integration by separating configuration, logic, and routing resources onto different layers. For example, the Rothko 3D FPGA architecture proposes a less symmetric architecture where alternating tiers ‘flow’ in opposite directions. The architecture and topology we chose for the 3D AFPGA is a symmetric and uniform extension from a similar 2D AFPGA it is just one of many possible architectures for 3D FPGAs. ![]() Alternative architectures may introduce programmable inter-tier switches directly into the logic block as a means of bypassing the global interconnect, reducing the inter-tier latency. In this topology, inter-tier communication must traverse at least two switches, one on the source tier and one on the destination tier. This decision keeps the switch overhead (in configuration logic, area, energy, and performance) to a minimum while accommodating 3D mappings. In the 3D prototype, we extended only one of the five switch points with inter-tier channels. Each switch box is linearly populated with five switch points. All five routing tracks only connect neighboring switch points no “long tracks” bypass switches. Each switch point contains two buffers to support two non-conflicting routes through each point. degrading the throughput of inter-process communication. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |