TSV Fault-Tolerance and Repair — 3D-IC Reliability

---

# 🏗️ TSV Fault-Tolerance and Repair Architectures in 3D-ICs
### *Through-Silicon Via Redundancy and Repair Strategies for Modern 3D Integration*

@luk036 👨‍💻 · 2026 📅

---

### 📋 Agenda

1. **Introduction to 3D-ICs** 🏗️ (3 min)
2. **TSV Fabrication Challenges** 🛠️ (3 min)
3. **Yield and Reliability Impact** 📊 (2 min)
4. **Redundant TSV Grouping** 🧩 (5 min)
5. **Clustered Defect Modeling** 📍 (3 min)
6. **BIST Architecture** 🔍 (4 min)
7. **TDMA Fault Tolerance** 🕒 (3 min)
8. **Switching vs. Shifting** 🔄 (5 min)
9. **Adaptive Fault Tolerance** 🛡️ (4 min)
10. **Graph-Based Optimization** 🔀 (5 min)
11. **Delay and Area Trade-offs** ⏱️ (3 min)
12. **Experimental Case Study** 🧪 (3 min)
13. **Conclusion** 🏁 (2 min)

> 📈 **Progression:** Understand → Detect → Repair → Optimize → Verify

---

### 🔬 Introduction to 3D-ICs

*   **Definition:** Vertically stacking multiple silicon dies connected by **Through-Silicon Vias (TSVs)**.
*   **Key Advantages:**
    *   ⚡ **Reduced Interconnect Delay:** Average wire-length varies with layers:
        $$\color{blue} L_{avg} \propto \sqrt{N_{layers}}$$.
    *   🔋 **Lower Power Dissipation:** Shorter vertical paths replace long horizontal wires.
    *   🚀 **High Bandwidth:** Massive parallel vertical connections.
    *   🤝 **Heterogeneous Integration:** Stacking logic, memory, analog, and sensors.

---

### 🛠️ The TSV Fabrication Challenge

TSV manufacturing is a complex physical process highly susceptible to failures:

*   **Stages:** Via etching, filling, thinning, and thermal bonding.
*   **Defect Modes:**
    *   ❌ **Open Defects:** Insufficient filling, voids, or micro-cracks.
    *   🔌 **Short Defects:** Substrate shorting or leakage.
    *   📏 **Geometric Defects:** Misalignment between dies or uneven TSV heights.
*   **Observation:** Even a single post-bond TSV failure can render an entire multi-die stack defective.

---

### 📊 Impact on Yield and Reliability

*   **Stacking Yield:** Cumulative product of die, bonding, and TSV yields:
    $$\color{blue} Y_{chip} = Y_{stack} \cdot \prod Y_{bonding(i)} \cdot \prod Y_{TSV(i)}$$
*   **Reliability Concerns:** Latent defects can emerge post-manufacturing due to thermal and physical stress during operation.
*   **The Solution:** Fault-tolerance architectures using **redundant (spare) TSVs** are essential to improve yield and lifespan.

---

### Part 1: Repair Fundamentals 🧩

---

### 🧩 Foundational Repair: Redundant TSV Grouping

The basic strategy partitions functional TSVs (f-TSVs) and redundant TSVs (s-TSVs) into independent groups.

.pull-left[
.mermaid[
<pre>
graph TD
    subgraph "TSV Group"
    A[Signal A] --> M1{MUX}
    B[Signal B] --> M1
    M1 --> T1((f-TSV 1))
    M1 --> T2((f-TSV 2))
    M1 --> S1((s-TSV))
    T1 -.-> D1[Die 2]
    T2 -.-> D1
    S1 -.-> D1
    end
    style T1 fill:#fff3e0,stroke:#e65100,stroke-width:3px
    style T2 fill:#fff3e0,stroke:#e65100,stroke-width:3px
    style S1 fill:#e3f2fd,stroke:#1565c0,stroke-width:3px
</pre>
]
]
.pull-right[
*   **Grouping Ratio:** Defined as $\color{blue} gr = N_{gr} : N_{gs}$ (Regular:Redundant).
*   **Logic:** A group is repairable if the number of defective TSVs (${\color{blue} x}$) is $\le$ the number of spare TSVs (${\color{blue} k}$).
]

---

### 📈 Yield Analysis: Independent Defects

Assuming each TSV fails independently with a uniform failure rate (${\color{blue} p}$):

*   **Binomial Distribution:** Probability of having exactly ${\color{blue} x}$ defective TSVs in a group of size $n$:
    $$\color{blue} P(X = x) = \binom{n}{x} p^x (1-p)^{n-x}$$.
*   **Group Yield:**
    $$\color{blue} Y_{group} = \sum_{x=0}^{k} \left[ \binom{n}{x} p^x (1-p)^{n-x} \right]$$.
*   **Total Yield:** $\color{blue} Y_{total} = (Y_{group})^t$, where $t$ is the number of groups.

---

### 📍 The Reality of Clustered Defects

In practice, TSV defects are spatially correlated due to bonding pressure variations or wafer roughness.

*   **Cluster Centers:** A single defect increases the likelihood of failures in its immediate vicinity.
*   **Defect Probability (${\color{blue} P[i]}$):** Inversely proportional to the distance (${\color{blue} d_c[i]}$) from a cluster center:
    $$\color{blue} P[i] = p \cdot \left( 1 + \left( \frac{1}{d_c[i]} \right)^\alpha \right)$$
    where $\color{blue} \alpha$ is the clustering coefficient.
*   **Insight:** Larger $\color{blue} \alpha$ values cause higher yield drops, especially in designs with high TSV density.

---

### Part 2: Test & Detection 🔍

---

### 🔍 Built-In Self-Test (BIST) Architecture

BIST enables chips to autonomously detect and record TSV failures.

.mermaid[
<pre>
graph TD
    subgraph "BIST Architecture"
        BC["BIST Controller"] --> PG["Pattern Generator"]
        PG --> TSV["TSV Array"]
        TSV --> RA["Response Analyzer"]
        RA --> FR["Fuse Register"]
    end
    style BC fill:#c8e6c9,stroke:#2e7d32,color:#2e3440,stroke-width:3px
    style PG fill:#e3f2fd,stroke:#1565c0,color:#2e3440,stroke-width:3px
    style TSV fill:#fff3e0,stroke:#e65100,color:#2e3440,stroke-width:3px
    style RA fill:#f3e5f5,stroke:#7b1fa2,color:#2e3440,stroke-width:3px
    style FR fill:#ffcdd2,stroke:#c62828,color:#2e3440,stroke-width:3px
</pre>
]

*   **Grouping Concept:** TSVs are divided into **Read Groups** (Slave to Master) and **Write Groups** (Master to Slave).
*   **Test Methodology:**
    1.  Apply test patterns (vectors) via the BIST controller.
    2.  Compare results on the master die using loopback paths for write groups.
    3.  Store "Repair Signatures" in **Fuse Registers**.
*   **Gain:** BIST grouping can reduce test time by approximately **17.5%**.

---

### 🕒 TDMA-Based Fault Tolerance

A cost-effective technique using Time Division Multiple Access to detect and bypass faulty links.

*   **Three Modules:** TDMA, Testing, and Routing.
*   **Sequential Testing:** Signals utilize specific time slots; the testing module identifies defects via delay tests (detecting voids/delamination) or short-to-substrate tests.
*   **Control Flow:**
    *   If **Testresult = 1**, the NMOS path of the defective TSV is cut off.
    *   Signal is re-routed through a defect-free TSV in a subsequent clock cycle.
*   **Overhead:** Minimal area cost, though it introduces serialized delay.

---

### Part 3: Repair Architectures 🛡️

---

### 🔄 Rerouting: Switching vs. Shifting

Different hardware structures balance repair rate and area overhead.

---

### 🛡️ Adaptive Fault Tolerance (AFTS)

AFTS adaptively determines the number of tolerant faults (${\color{blue} K}$) based on f-TSV and s-TSV distributions.

*   **Constraint Violation:** Previous works assumed fixed ${\color{blue} K}$ and that any s-TSV could replace any f-TSV.
*   **AFTS Advantage:** Handles non-uniform layouts where f-TSVs have limited candidate s-TSVs.
*   **Optimization:** Uses **Integer Linear Programming (ILP)** to minimize MUX delay and the number of s-TSVs used.
*   **Result:** Reduces spare TSV count by up to **48%** compared to fixed-ratio methods.

---

### Part 4: Graph-Based Optimization 🔀

---

### 🔀 Solving Repair via Max Flow

Repair problems are modeled using graph theory, specifically vertex-disjoint paths.

*   **Theorem (Menger):** The max number of vertex-disjoint paths equals the minimum disconnecting vertex set.
*   **Node Splitting Transformation:** To solve using standard edge-flow algorithms, each node ${\color{brown} u}$ is split into ${\color{brown} u}$ (input) and ${\color{brown} u'}$ (output) connected by a zero-cost edge.

.mermaid[
<pre>
graph LR
    subgraph "Split Node"
    U((u_in)) -- "Capacity=1, Cost=0" --> Up((u_out))
    end
    Source -- "Capacity=1" --> U
    Up -- "Replaceable Relation" --> V((v_in))
    style U fill:#e3f2fd,stroke:#1565c0,stroke-width:3px
    style Up fill:#ffccbc,stroke:#bf360c,stroke-width:3px
    style Source fill:#e3f2fd,stroke:#1565c0,stroke-width:3px
    style V fill:#e3f2fd,stroke:#1565c0,stroke-width:3px
</pre>
]

---

### ⚖️ Integer Linear Programming (ILP) Formulation

ILP is used to find the optimal repair structure with minimum delay overhead.

*   **Objective:** Minimize the maximum indegree ($\color{blue} \lambda$) of all vertices to limit MUX input port counts:
    $$\color{blue} \min \max_{u \in V} \{ indegree(u) \}$$.
*   **Binary Variables:** $\color{blue} x_{vu}^{(s,t)} = 1$ if a path from f-TSV $s$ to s-TSV $t$ passes through edge $(u, v)$.
*   **Constraint:** Ensures paths are edge-disjoint (vertex-disjoint in original graph).

---

### 🏎️ MCMF Heuristic for Large Groups

For massive TSV arrays, ILP becomes computationally expensive; **Min-Cost Max-Flow (MCMF)** heuristics provide a faster alternative.

*   **Process:** Solve edge-disjoint paths for f-TSVs one by one.
*   **Cost Function:**
    *   $\color{blue} Cost = 0$ for already used edges.
    *   $\color{blue} Cost = C$ for introducing a new spare TSV to the repair chain.
*   **Goal:** Re-use existing TSV connections to minimize total MUX ports and wirelength.

---

### Part 5: Practical Considerations ⚖️

---

### ⏱️ Delay and Area Overhead Trade-offs

Repairing a TSV is not "free"—it introduces signal degradation.

*   **Wire Delay:** Signals must travel longer distances to reach redundant TSVs.
    *   Example: A 400µm reroute adds ~200ps delay.
*   **Gate Delay:** Larger MUXs required for high repair rates increase path latency.
    *   Example: A 14:1 MUX adds ~340ps.
*   **Critical Timing:** Timing-aware placement ensures the most critical signals are least affected by rerouting.

---

### 🧪 Experimental Case Study: HBM Yield

High-Bandwidth Memory (HBM) stacking is particularly sensitive to TSV defects.

*   **Observation:** Standard repair schemes fail under the heavy clustering typical of HBM.
*   **LCTRA Performance:** Research indicates BIST repair can improve HBM yield by **12.5%**.
*   **Lifespan Enhancement:** Built-in self-repair (BISR) can monitor and bypass interconnects that degrade during the chip's operating life.

---

### 🏁 Conclusion

*   🏗️ **TSVs are critical but fragile** components of 3D integration.
*   📉 **Yield loss** is the primary barrier to commercial 3D IC adoption.
*   🛡️ **Modern repair architectures** combine BIST, redundant grouping, and advanced graph-based optimization to reach **~100% yield**.
*   🔮 **Future Work:** Development of temperature-aware recovery methods and accurate aging models to further enhance reliability.

---

# 🙏 Thank You!
### *Questions?*

**Slides:** https://luk036.github.io/net_optim/spare-tsv-remark.html

> 💡 **"Reliability is not optional — it's the foundation of 3D integration."**