die stacking

```svg 3D IC: go vertical — bond two wafers, or build the second tier in placeParallel 3D bonds finished wafers via TSV or Cu-Cu; monolithic 3D grows a new transistor tier on top, in place1 · Two ways to go verticalParallel — bonddie 2die 1TSV / Cu-Cu bondMonolithic — in placetier 2tier 1inter-tier via (nm)Parallel bonds two finishedwafers; monolithic grows tier 2directly on tier 1 — no bond.Inter-tier vias are far denserthan any bonded connection —that is the whole point.2 · Vertical pitch laddermicrobump~30–40 µmTSV~5–10 µmCu-Cu hybrid bond< 1 µmmonolithic inter-tier via~50–100 nmFiner vertical pitch = finer 3Dpartitioning: whole chips, thenblocks, then individual gates.Bonding buys density; monolithicbuys another 100x on top.3 · Thermal budget & payoffsPayoffsShorter global wiresMemory directly over logicCFET: nFET over pFET, tiny cellsNew floorplans across tiersThe thermal wallTier-2 built cold (<~500 °C) sotier-1 devices surviveBuried tiers are hard to coolYield multiplies across tiersHeat removal and low-temp devicequality are the real limits.Bond or build-in-placeStack and bond two finished wafers,or grow a second transistor tiersequentially on the first.Density sets the payoffFrom TSV µm to monolithic-via nm,finer vertical pitch moves you fromchip-level to gate-level 3D.Heat & thermal budget biteBuried tiers are hard to cool, andmonolithic's top tier must be builtcold enough to spare the bottom. ``` **3D IC Integration and Die Stacking** encompasses the **technologies for vertically stacking multiple semiconductor dies and connecting them with through-silicon vias (TSVs), hybrid bonding, or other vertical interconnects** — creating three-dimensional integrated circuits that achieve higher bandwidth, lower power, greater heterogeneous integration density, and smaller footprint than equivalent 2D implementations. **3D Stacking Approaches:** ``` Packaging Hierarchy (increasing integration density): 2.5D: Dies side-by-side on silicon interposer (CoWoS, EMIB) Interconnect: RDL on interposer, 25-55μm bump pitch BW: 100s GB/s between dies Example: HBM stacks next to GPU on interposer 3D (TSV): Dies stacked vertically, connected by TSVs Interconnect: TSVs (~5-10μm diameter, ~50μm pitch) BW: TB/s (thousands of TSV connections) Example: HBM DRAM stacks (4-16 die) 3D (Hybrid Bond): Die-to-die or wafer-to-wafer Cu-Cu direct bonding Interconnect: sub-10μm pitch Cu pads BW: Multi-TB/s (millions of connections) Example: AMD V-Cache, Sony image sensors Monolithic 3D: Sequential transistor fabrication on same wafer Interconnect: Inter-layer vias at gate pitch (research stage — CFET is a form of this) ``` **TSV Technology:** | Parameter | Value | |-----------|-------| | TSV diameter | 5-10μm (fine), 20-50μm (coarse) | | TSV pitch | 20-50μm (fine), 100-200μm (coarse) | | TSV depth | 40-100μm (after die thinning) | | Aspect ratio | 5:1 to 10:1 | | Fill material | Electroplated copper | | Liner/barrier | SiO₂ isolation + TaN/Ta + Cu seed | | Resistance | <50mΩ per TSV | | Capacitance | ~30-50fF per TSV | | Process | Via-first, via-middle, or via-last | **Hybrid Bonding:** The most advanced D2D connection technology: ``` Process: 1. Prepare bonding surfaces: CMP Cu pads and SiO₂ dielectric Surface roughness: <0.5nm RMS Cu recess: 2-5nm below oxide surface 2. Surface activation: plasma treatment (N₂/O₂) Creates hydrophilic surface for bonding 3. Room-temperature oxide bonding: face-to-face alignment SiO₂-SiO₂ van der Waals bonding at room temperature Alignment accuracy: <200nm (W2W), <500nm (D2W) 4. Anneal at 200-400°C: Cu expands, Cu-Cu metallic bond forms Cu CTE (17ppm/°C) > SiO₂ CTE (0.5ppm/°C) → Cu pad pushes up and contacts opposing Cu pad Result: Simultaneous electrical + mechanical bond at <10μm pitch (10,000-1,000,000+ connections per mm²) ``` **Applications:** | Application | Technology | Example | |------------|-----------|--------| | HBM memory | TSV stacking (8-16 die) | SK Hynix HBM3E | | Cache stacking | Hybrid bonding (D2W) | AMD V-Cache (3D V-Cache) | | Image sensors | Hybrid bonding (W2W) | Sony IMX stacked CIS | | AI accelerators | 2.5D + 3D hybrid | NVIDIA B200, AMD MI300 | | FPGA | Die stacking | Intel FPGA (Agilex) | **Design Challenges:** - **Thermal**: Bottom die in stack is farthest from heat sink. Power density limits: ~2W/mm² total for air-cooled stacked dies. - **Testing**: KGD required before bonding (no rework possible after hybrid bonding). - **Stress**: CTE mismatch between stacked dies causes warpage and stress on TSVs/bonds. - **EDA**: 3D physical design tools must handle multi-die floorplanning, inter-die routing, and thermal co-optimization. **3D IC integration is the primary scaling vector for the post-Moore era** — when lateral transistor scaling can no longer provide sufficient performance gains, vertical integration enables continued improvement in bandwidth density, functional density, and heterogeneous integration, making 3D stacking the defining technology trend in advanced semiconductor packaging.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account