MIN:TYP:MAX and Delay Selection

Chapter 4 · Section 4.4 · SDF Annotation

Project thread — 4.3 applied the counter's SDF; this lesson decides which of each triplet's three values the counter run uses. 4.6 picks one deliberately.

1. Why Should I Learn This?

Every SDF delay is three numbers; your run uses one. Pick the wrong one and your "timed" run stresses the wrong thing.

• MAX = slow → setup pressure; MIN = fast → hold pressure; TYP = nominal.
• Selection is an explicit option — and tool defaults vary.
• The triplet is within one SDF; corners are separate SDFs (3.4) — related but different axes.

This follows the annotation flow (4.3) and precedes debugging (4.5) and the counter capstone (4.6).

2. Real Silicon Story — the setup margin that TYP hid

A team ran SDF-annotated GLS at the typical delay value and saw comfortable timing. STA (at the slow corner) later flagged a setup violation.

The GLS run used TYP — the nominal delay — so logic arrived with nominal, not worst-case, delay. MAX (the slow value in the triplet) would have pushed arrival later and made setup pressure visible in the functional behaviour. Running TYP quietly understated the delay that setup cares about.

Lesson: the triplet value you select changes what your run can show. Setup is a MAX (slow) concern; running TYP or a tool default can hide it.

3. Concept — the triplet and how it's selected

The triplet: each SDF delay is (MIN:TYP:MAX):

• MIN — fastest / shortest delay.
• TYP — typical / nominal.
• MAX — slowest / longest delay.

Selection (at annotation):

• Run option: +mindelays / +typdelays / +maxdelays, or an SDF command file.
• The simulator applies one value from each triplet to the arcs.
• Defaults vary by tool — never assume; state it explicitly.

Which value for which check:

• MAX (slow) → stresses setup (logic must be fast enough).
• MIN (fast) → stresses hold (data must hold past the edge).
• TYP → nominal behaviour, not a worst case.

Triplet vs corners (important):

• The triplet is min/typ/max within one characterised SDF (one corner).
• Corners (3.4) are separate SDFs from separate .libs.
• Both shift delay, but they are different axes — and neither GLS selection replaces all-corner STA (0.3).

4. Mental Model — three speed estimates for the same trip

Each delay triplet is three estimates of the same signal's travel time.

• MIN = the optimistic estimate (everything fast) → relevant to hold (things arriving too soon).
• MAX = the pessimistic estimate (everything slow) → relevant to setup (things arriving too late).
• TYP = the average day → not the worst case for either.

Selecting a value is choosing which estimate your run believes. A worst-case setup question needs the pessimistic (MAX) estimate — the average won't reveal the tight margin.

5. Working Example — the triplet, selection, and a min-vs-max waveform

The same arc, one triplet, two selections:

# SDF triplet — REPRESENTATIVE. One delay, three values: MIN:TYP:MAX
(IOPATH CK Q (0.08:0.11:0.15))
#            MIN  TYP  MAX
#   +mindelays -> 0.08  (fast; hold-flavoured)
#   +typdelays -> 0.11  (nominal)
#   +maxdelays -> 0.15  (slow; setup-flavoured)

Practical context (representative, tool-neutral):

# Select the triplet value explicitly (tool-neutral idea):
#   setup-flavoured GLS:  annotate with MAX   (+maxdelays or command-file MAX)
#   hold-flavoured GLS:   annotate with MIN   (+mindelays)
#   nominal functional:   TYP
# DO NOT rely on the tool default (varies). State the selection in the run log.
# Reminder: triplet != corners. Different SDFs per corner (3.4); STA covers all corners.

MIN vs MAX arrival at a capture edge, as a real waveform:

6. Debugging Session — a comfortable TYP run that hid a setup violation

7. Common Mistakes

• Trusting the tool default. Defaults vary — select explicitly.
• Running TYP for a setup question. Setup needs MAX; TYP understates delay.
• Running MAX for a hold question. Hold needs MIN.
• Confusing the triplet with corners. Triplet is within one SDF; corners are separate SDFs (3.4).
• Treating one selection as timing closure. STA across all corners is signoff (0.3).

8. Industry Best Practices

• Select MIN/TYP/MAX to match the check (MAX→setup, MIN→hold).
• State the selection explicitly and log it (reproducibility).
• Keep triplet and corner axes distinct — both matter, differently.
• Use an SDF command file for corner + selection + scaling.
• Defer closure to STA across all corners; use selection to stress GLS functionally.

Senior Engineer Thinking

• Beginner: "I annotated the SDF, so the timing is real."
• Senior: "Which value of the triplet did I apply — and was it the tool default? Setup is MAX, hold is MIN. And this is one selection of one SDF; all-corner closure is STA."

The senior chooses the triplet value deliberately and never confuses it with corner coverage.

Silicon Impact

Silicon delays span the whole MIN–MAX range at every corner. Verify at TYP (or an unknown default) and call it done, and a setup margin (MAX) or hold margin (MIN) can be understated in your functional runs — letting a real violation slip toward silicon as a marginal, condition-dependent failure (0.3). The triplet exists precisely because one delay value never represents the spread. GLS min/typ/max selection lets a run stress the right extreme functionally; STA across all corners and both extremes is what proves the margins.

Engineering Checklist

• Selected the triplet value to match the check (MAX→setup, MIN→hold).
• Made the selection explicit (not the tool default) and logged it.
• Kept triplet selection distinct from corner selection (3.4).
• Used an SDF command file for corner + selection where needed.
• Relied on STA (all corners, both extremes) for closure.

Try Yourself

1. Annotate the representative (IOPATH CK Q (0.08:0.11:0.15)) with +maxdelays and check a tight path against the clock — setup pressure appears (late arrival).
2. Observe: at MAX the result arrives latest.
3. Change: re-annotate with +mindelays (0.08) and re-run.
4. Expect: the path arrives earliest — setup pressure gone, but hold is now the concern. Confirm TYP sits between, and that no single selection shows both.

Any free Verilog simulator supports +mindelays/+typdelays/+maxdelays. Real triplets come from characterisation, but selection behaves identically. No paid tool required.

Interview Perspective

• Weak: "The SDF has one delay per arc."
• Good: "Each SDF delay is a MIN:TYP:MAX triplet; the simulator picks one at annotation."
• Senior: "MAX is the slow value that stresses setup, MIN the fast value that stresses hold, TYP nominal. I select explicitly — defaults vary — and I keep the triplet distinct from corners: it's min/typ/max within one SDF, while corners are separate SDFs. All-corner closure is STA, not any single selection."

9. Interview / Review Questions

10. Key Takeaways

• Each SDF delay is a MIN:TYP:MAX triplet — MIN fastest, TYP nominal, MAX slowest — and the simulator applies one at annotation.
• Select via a run option (+mindelays/+typdelays/+maxdelays) or an SDF command file; tool defaults vary, so choose explicitly.
• MAX (slow) stresses setup; MIN (fast) stresses hold; TYP is nominal — the same intuition as corners.
• The triplet is min/typ/max within one SDF (one corner); corners are separate SDFs (3.4) — a different axis.
• No single selection is timing closure — STA across all corners and both extremes is the signoff (0.3). Next: 4.5 — debugging failed & partial SDF annotation.

Quick Revision

Each SDF delay = MIN:TYP:MAX triplet; simulator applies one (+min/+typ/+maxdelays or command file; defaults vary). MAX→setup, MIN→hold, TYP→nominal. Triplet is within one SDF; corners are separate SDFs (3.4) — different axis. No single selection = closure; STA (all corners) is signoff. Next: 4.5 — debugging failed & partial annotation.