standard cell characterization

Standard Cell Library Characterization is the process of measuring and modeling static/dynamic behavior of logic cells across voltage/temperature/process corners, producing Liberty (.lib) files that enable accurate timing closure and power analysis in SoC design.

Liberty (.lib) Format and Structure

• Liberty File Format: Text-based specification of cell timing/power characteristics. Defines pins, functions, timing arcs, power tables in human-readable/machine-parseable form.
• Cell Definition: Each cell (NAND2, NOR3, flip-flop) contains pin descriptions (input/output), function (Boolean logic), timing models, power dissipation.
• Pin Declaration: Input/output pins specified with direction, capacitance, rise/fall slew rate transitions. Internal pins for special functions (clock, reset).
• Timing Arc: Connection from one pin to another with delay/slew characterization. Example: NAND2 has A→Y, B→Y delay arcs; flip-flop has D→Q, CLK→Q, SET→Q arcs.

NLDM and CCS Timing Models

• NLDM (Non-Linear Delay Model): Delay and transition time tables indexed by input slew rate and output load capacitance. Cubic polynomial interpolation between table values.
• Delay Formula: Delay = f(input_slew, output_load). NLDM provides 2D lookup tables (slew × load). Typical table: 5×5 or 7×7 (25-49 characterization points per arc).
• CCS (Composite Current Source): Current-based timing model. Cell output modeled as time-varying current source. Accuracy > NLDM for complex waveform scenarios (glitch, crosstalk).
• CCS Advantages: Captures frequency-dependent behavior, crosstalk noise impact, multi-input switching. Enables better STA accuracy but ~5x larger Liberty files vs NLDM.

Cell Delay and Propagation Arcs

• Propagation Delay (Tpd): Time from input transition 50% to output transition 50%. Monotonically increases with load capacitance and input slew rate.
• Slew Propagation: Output slew (rise time, fall time) characterized similarly. Impacts fanout gate delays (higher slew = longer downstream delays).
• Delay Dependencies: Temperature effect (negative temperature coefficient: faster at low T), supply voltage (lower voltage → higher delay), process (Vth variation → delay variation).
• Multi-Input Cells: Complex cells like muxes, adders have multiple delay arcs (each input → each output). NAND8 has 8 delay paths; characterization combinatorial explosion addressed via clustering/approximation.

Setup/Hold and Clock-to-Q Timing Arcs

• Setup Time: Minimum time data must be stable before clock transition. Library specifies setup for all data pins (D, preset, clear) vs clock.
• Hold Time: Minimum time data must remain stable after clock transition. Hold violations more serious than setup (can't pipeline out of hold).
• Recovery/Removal Times: For asynchronous inputs (reset, preset). Recovery = minimum time reset must release before clock. Removal = hold-like constraint on reset relative to clock.
• Clock-to-Q Delay: Delay from clock edge to output switching. Highly load-dependent. Critical for timing budgeting in datapaths.

PVT Characterization Corners

• Process Variation: Fast (Vth low, gate oxides thin), slow (opposite), typical corners. SPICE simulations at nominal/extreme process parameters.
• Voltage Variation: Nominal (1.2V), high (1.35V), low (1.05V). Simulations re-run at each supply voltage. Voltage scaling dramatically affects delay.
• Temperature Variation: Nominal (25°C), high (85°C or 125°C), low (0°C or -40°C). Temperature affects Vth (negative coefficient) and carrier mobility (positive).
• Typical Characterization: 3×3×3 (process × voltage × temperature) = 27 Liberty files. High-end libraries may include additional intermediate points.

Statistical (SSTA) Liberty Extensions

• Statistical Variation Modeling: SSTA acknowledges not all corners equally likely. Process variation follows normal distribution; characterize sigma (σ).
• Sigma Tables: Liberty extended with statistical parameters. Cell delay μ (mean) and σ (standard deviation) of delay distribution vs PVT corners.
• Parametric Variation: Cell delay model includes random variables (Vth mismatch, length variation) beyond fixed corners. Enables better yield prediction.
• Correlation: Delay variations across multiple cells correlated (spatially correlated process effects). Statistical models capture correlation reducing pessimism in STA.

Characterization Methodology

• Spice Simulation Setup: SPICE netlist of cell with transistor-level models (BSIM4, BSIM6). Stimulus: input ramp (multiple slew rates), load capacitor varied (5-500fF typical).
• Measurement Points: Simulations measure delay, slew, power (switching + leakage) for each (slew, load, corner) combination.
• Table Generation: Measured data interpolated to regular grid. Polynomial fitting reduces sensitivity to simulation noise.
• Liberty Generation: Automated tools (Cadence Liberate, Synopsys Characterizer) convert SPICE results to Liberty file with formatting and verification.