statistical static timing analysis (ssta)

**Statistical Static Timing Analysis (SSTA)** is a timing verification methodology that treats **all delays as probability distributions** rather than fixed numbers — computing the **statistical distribution of path delays** to determine the probability that timing constraints are met, rather than simply reporting pass/fail against worst-case deterministic values. **SSTA vs. Deterministic STA** - **Deterministic STA**: Each cell/wire delay is a single number (at a given corner). The path delay is the sum. Timing either passes or fails. Uses OCV/AOCV/POCV derates to add margin for variation. - **SSTA**: Each delay is a distribution (mean + variance + correlations). The path delay is a resulting distribution. The output is a **probability of timing failure** rather than a binary pass/fail. **Why SSTA Is Needed** - At advanced nodes, process variation is large relative to the timing margin — **deterministic worst-case analysis becomes overly pessimistic**. - In deterministic STA, taking worst-case at every step compounds pessimism — the probability that EVERY cell is simultaneously at worst case is vanishingly small. - SSTA models variations statistically and accounts for correlations — producing a **realistic estimate** of timing yield. **How SSTA Works** - **Variation Sources**: Identify all sources of variation — inter-die (global corner), intra-die systematic (spatial gradients), intra-die random (device-level). - **Delay Models**: Each gate and wire delay is expressed as: $$d_i = d_{i,nom} + \sum_j a_{ij} \cdot \Delta p_j + r_i$$ Where $d_{i,nom}$ is the nominal delay, $a_{ij}$ are sensitivities to global/systematic variation sources $\Delta p_j$, and $r_i$ is the random component. - **Statistical Operations**: Replace deterministic addition and max operations with their statistical equivalents: - **Statistical ADD**: Sum of Gaussians is Gaussian — means add, variances add (for uncorrelated) or combine with correlation. - **Statistical MAX**: The maximum of two Gaussian variables is a new distribution — computed using Clark's method or moment matching. - **Output**: The arrival time at each node is a distribution. Slack is a distribution. The tool reports the **probability that slack is negative** (timing failure probability). **SSTA Outputs** - **Mean Slack**: The average timing margin. - **Sigma of Slack**: The variation in timing margin. - **Timing Yield**: Probability that timing is met — e.g., 99.73% (3σ), 99.99% (4σ). - **Parametric Yield**: Probability that the chip meets its frequency target. **SSTA Challenges** - **Computational Complexity**: Statistical MAX is the bottleneck — approximate methods are needed to keep runtime tractable. - **Correlation Modeling**: Must accurately model which variations are correlated (spatial, global) vs. uncorrelated (random). Incorrect correlation assumptions lead to inaccurate results. - **Adoption**: Industry has largely adopted **POCV within deterministic STA** as a practical approximation to full SSTA — POCV captures most of the benefit with less disruption to the existing flow. SSTA is the **theoretical gold standard** for variation-aware timing analysis — while full SSTA is not universally deployed, its concepts underpin AOCV and POCV, which are the practical state-of-the-art.