Compare NLDM and CCS for multivoltage scaling - which is the right choice and why?
From PDVerse STA Mentor Guide, part of the pdVerse Mentor Guide
Short Answer
CCS (composite current source) is the recommended Liberty model for scaling library groups: it represents the cell driver as a current source and captures nonlinear voltage dependence, whereas NLDM's (nonlinear delay model) delay and slew tables carry no real voltage dependence and drift from SPICE across a wide voltage sweep.
Technical Explanation
NLDM (nonlinear delay model) describes a cell with delay and output-slew tables indexed by input slew and output load, but those tables carry no real dependence on supply voltage of their own.
- Where NLDM holds up: within a narrow voltage band, the fixed-shape assumption behind the tables is close enough to what SPICE reports.
- Where NLDM breaks down: stretch the same tables across a wide voltage sweep and they progressively diverge from SPICE, because the shape of the driving waveform, not just its scale, changes with supply voltage — something a table with no voltage axis can't represent.
- What CCS does differently: CCS (composite current source) models the driver as a current source rather than a fixed delay lookup, so it reproduces the driver's nonlinear dependence on supply voltage much more faithfully.
- Why CCS is recommended for scaling groups specifically: multivoltage analysis is, by definition, running the same cells at several different voltages and interpolating between characterized points — exactly the regime where NLDM's error is largest and CCS's extra accuracy earns its cost.
- The practical consequence: the accuracy gap between the two models is smallest where scaling isn't needed and largest where it is. Treating NLDM as good enough for a DVFS (dynamic voltage and frequency scaling) design's scaling libraries imports an error that grows with the very voltage range the scaling was built to cover.
Common Mistake
The Trap: reusing NLDM libraries for a DVFS block because they already exist and are cheaper to characterize.
- A team building voltage-scaling library groups assumes NLDM's error is roughly the same everywhere and reuses it to save characterization time.
- The error is smallest at the nominal voltage the NLDM tables were centered on and grows toward the low end of the DVFS range, meaning the corner that matters most for power savings is exactly where the timing model is least trustworthy.
Follow-up Question & Model Response
If CCS costs more to characterize and simulate, is there ever a legitimate case for staying with NLDM in a multivoltage flow?
Candidate Model Response: Yes, for domains that operate at a single fixed voltage with no scaling, where NLDM's narrow-band accuracy is never stressed and its lower characterization and runtime cost is a real advantage. The decision should be scoped per voltage domain rather than applied uniformly across the whole chip — a domain with a wide DVFS range needs CCS libraries, while a domain pinned to one always-on rail can reasonably stay on NLDM if runtime is a genuine constraint. Mixing the two within one signoff run is normal and expected in a multivoltage design.
Practical Example
A DVFS core domain scales from 1.0V down to 0.6V. NLDM-based timing at 0.6V shows a buffer cell delay of 42 ps, while SPICE reports 58 ps at that same corner, a 28% underestimate that widens as voltage drops further. Switching that domain's library group to CCS brings the reported delay to 56 ps, within 4% of SPICE. A neighboring always-on I/O domain fixed at 1.0V keeps its existing NLDM libraries, since it never scales and the narrow-band NLDM accuracy holds.
Complete STA Handbook
Master Signoff-Ready Static Timing Analysis
Get the complete 10-chapter STA handbook covering setup/hold margins, clock modeling, OCV/POCV, crosstalk noise, and PrimeTime closure.
Offline PDF Bundle
Want all 1109 questions offline?
Get the complete 4-book PDF bundle (PnR, STA, MMMC, Low Power) with a clickable table of contents - no ads, no internet needed.
Continue practising