advanced packaging 3D IC

**Advanced Semiconductor Packaging** is **the collection of technologies that integrate multiple dies, chiplets, and passive components into compact, high-performance packages using 2.5D/3D stacking, hybrid bonding, and fan-out redistribution — enabling continued system-level scaling when transistor scaling alone cannot deliver the required performance, bandwidth, and energy efficiency improvements**. **2.5D Integration (Interposer-Based):** - **Silicon Interposer**: passive silicon substrate with through-silicon vias (TSVs) and fine-pitch redistribution layers (RDL); connects multiple chiplets with <10 μm bump pitch; TSMC CoWoS, Intel EMIB are leading platforms - **Bandwidth**: silicon interposer provides >1 TB/s aggregate bandwidth between chiplets; HBM (High Bandwidth Memory) stacks connected via interposer deliver 460-1200 GB/s per stack; critical for AI accelerators (NVIDIA H100 uses CoWoS with 5 HBM3 stacks) - **Organic Interposer**: lower cost alternative using organic substrate with embedded silicon bridge dies (Intel EMIB); bridge die provides fine-pitch connectivity only where needed; reduces cost vs full silicon interposer - **Thermal Challenges**: multiple high-power chiplets on shared substrate create thermal hotspots; thermal interface materials, heat spreaders, and liquid cooling required for >500W packages **3D Integration (Die Stacking):** - **Hybrid Bonding (Cu-Cu)**: direct copper-to-copper bonding at <1 μm pitch without solder bumps; oxide-oxide bonding provides mechanical adhesion; enables >10,000 connections per mm² — 100× denser than micro-bumps - **TSV Technology**: through-silicon vias (5-10 μm diameter, 50-100 μm depth) provide vertical electrical connections; via-first, via-middle, and via-last process flows depending on integration point; TSV capacitance ~30-50 fF limits high-speed signaling - **Wafer-on-Wafer (WoW)**: bond complete wafers face-to-face before dicing; highest throughput and alignment accuracy (<200 nm overlay); TSMC SoIC uses WoW for logic-on-logic stacking - **Die-on-Wafer (DoW)**: place known-good dies on wafer; enables heterogeneous integration of dies from different wafer sizes and process nodes; lower throughput but higher yield than WoW **Fan-Out Packaging:** - **Fan-Out Wafer-Level Packaging (FOWLP)**: dies embedded in molding compound with RDL extending connections beyond die edge; eliminates package substrate; TSMC InFO used for Apple A-series processors - **Fan-Out Panel-Level Packaging (FOPLP)**: processing on large rectangular panels (510×515 mm) instead of round wafers; higher throughput and lower cost per unit; Samsung, ASE developing FOPLP - **Multi-Die Fan-Out**: multiple chiplets embedded in single fan-out package; RDL provides die-to-die connectivity; cost-effective alternative to silicon interposer for moderate bandwidth requirements - **RDL Pitch**: advanced fan-out achieves 2-5 μm line/space in RDL; enables high-density routing comparable to silicon interposer at lower cost **Industry Ecosystem:** - **OSAT (Outsourced Assembly and Test)**: ASE, Amkor, JCET provide packaging services; increasingly investing in advanced packaging capabilities previously exclusive to foundries - **Foundry Packaging**: TSMC (CoWoS, InFO, SoIC), Intel (EMIB, Foveros), Samsung (I-Cube, X-Cube) vertically integrating packaging with wafer fabrication - **Standards**: UCIe (Universal Chiplet Interconnect Express) standardizes die-to-die interfaces; enables multi-vendor chiplet ecosystems; bandwidth up to 1.3 TB/s per mm of edge - **Market Growth**: advanced packaging market exceeding $50B by 2028; driven by AI accelerator demand (each NVIDIA H100/B200 requires CoWoS packaging); capacity constraints driving massive investment Advanced semiconductor packaging is **the critical enabler of continued system performance scaling in the post-Moore era — by integrating heterogeneous chiplets through increasingly sophisticated interconnect technologies, packaging has evolved from a commodity back-end process to the strategic differentiator defining next-generation computing architectures**.