Difference between AOCV , OCV and POCV - Timing / STA

Beyond the Derate: OCV vs AOCV vs POCV in STA

As process nodes shrink below 28nm, On-Chip Variation (OCV) becomes a primary driver of timing closure failures. This article explains the three variation methodologies — OCV, AOCV, and POCV — with a worked setup/hold example you can follow along.

📑 Contents

1. Why Do We Need Variation Modeling?
2. OCV — The Blunt Instrument
3. AOCV — The Statistical Standard
4. POCV — The Precision Tool
5. Comparison Table
6. Worked Example: Setup & Hold
7. Practical Tips

1. Why Do We Need Variation Modeling?

Variation in CMOS processes stems from two sources:

• Global Variation: Systematic differences across the wafer (oxidation thickness, lithography focus). Affects all transistors similarly.
• Local Variation: Random differences between neighboring transistors (Random Dopant Fluctuations, Line Edge Roughness). Spatially uncorrelated.

The goal: Capture the worst-case delay caused by variation — without adding excessive guardband that wastes area and power.

2. OCV — The "Blunt Instrument"

Standard OCV applies a global, fixed margin to all timing paths via delay derating:

T_final = T_nominal × (1 + Derate)

Critical flaw — correlation ignorance: OCV treats a 5-cell path inside a small block exactly the same as a 100-cell path crossing half the die. Short paths get far more margin than statistically needed, causing false violations and over-design.

💡 When to use: Early exploration, mature nodes (>65nm), or when statistical foundry data is unavailable.

3. AOCV — The Statistical Standard

AOCV uses foundry sigma (σ) tables and recognizes that local variation accumulates with path length. The margin scales with the number of stages:

Margin = σ_global + σ_local × √N

The √N term is key: margin grows with path length, but at a decreasing rate. Short paths get small margins; long paths get proportionally larger ones.

• ✅ Industry standard for 28nm → 7nm
• ✅ Significantly de-pessimized vs. OCV
• ⚠️ Path-based only — blind to physical location on the die

4. POCV — The Precision Tool

POCV moves from path-level to cell-level analysis. Each cell gets a unique sigma based on its exact location, local environment, and spatial correlation:

Margin = Σ ( w_i × σ_i )

Where w_i is a correlation weight based on cell positions. This captures local spatial variation that AOCV cannot see.

• ✅ Highest accuracy, least pessimism — essential at 5nm/3nm/2nm
• ⚠️ Complex setup, slower runtime, requires detailed foundry parametric files

5. Comparison Table

Feature	OCV	AOCV	POCV
Granularity	Global (whole chip)	Path-based (statistical)	Cell-based (parametric)
Local variation	❌ Not captured	⚠️ Approximated (√N)	✅ Fully captured
Pessimism	High	Medium	Low
Runtime	Fast	Moderate	Slow
Typical node	>65nm	28nm–7nm	7nm and below
Data required	User derates	Foundry sigma tables	Per-cell sigma maps

6. Worked Example: Setup & Hold

Setup (long path): data arrives too late → variation must worsen the data path. Hold (short path): data arrives too early → variation must shrink the data path.

Setup Analysis — Long Path (D_data = 950ps, T_clk = 1000ps, T_setup = 50ps)

Method	Margin	Final Delay	Slack	Result
Nominal	0	950ps	0ps	Critical
OCV (3%)	+28.5ps	978.5ps	−28.5ps	FAIL
AOCV (3σ)	+30ps	980ps	−30ps	FAIL
POCV	+25ps	975ps	−25ps	FAIL (least pessimistic)

Hold Analysis — Short Path (D_data = 60ps, T_hold = 55ps)

Method	Margin	Final Delay	Slack	Result
Nominal	0	60ps	5ps	Pass
OCV (3%)	−1.8ps	58.2ps	3.2ps	Pass
AOCV (3σ)	−6ps	54ps	−1ps	FAIL
POCV	−4ps	56ps	1ps	Pass (tight)

⚠️ Interesting insight: On this short hold path, AOCV is more pessimistic than OCV because the statistical 3σ margin exceeds the fixed 3% derate. POCV fixes this using local sigma data — the choice of methodology directly changes pass/fail results!

7. Practical Tips

• Never sign off with plain OCV below 28nm — it is too coarse.
• Monitor AOCV hold fails: many new hold violations after switching from OCV? Check your sigma tables and correlation length.
• POCV saves power: fewer buffers, smaller hold-fix delay cells, tighter clock trees.
• A "fail" in OCV may be a "pass" in AOCV — always know which methodology your report uses.

Bottom line: OCV = simple but pessimistic. AOCV = balanced statistical approach. POCV = maximum precision for advanced nodes. Choose based on your node, data availability, and accuracy needs.

⚡

Enjoyed this article?

For more tutorials on STA, RTL Design, CDC, and VLSI — visit my blog:

Visit RTL Digital Design →