Back to Viridian

SEMICONDUCTOR & EDA RESEARCH / SILEX

Semiconductor Physical Design & EDA Research

Mapped structure, multi-corner timing and evidence-driven optimisation qualification.

Retained SILEX engineering research · Not production qualified

Engineering motivation

Physical-design optimisation is a constrained engineering problem. Reducing delay on one path can change load, fanout, placement or cell area; an intervention that helps setup may degrade hold. An isolated improvement cannot establish overall implementation superiority.

Viridian’s SILEX research investigates how technical constraints, physical checkpoint identity and reproducible comparisons can guide optimisation decisions. The work operates around an OpenROAD/OpenSTA implementation flow. It is ongoing engineering research, not a new signoff tool or an established replacement for commercial EDA software.

Methodology and qualification

Mapped structure and physical constraints

Technology mapping determines the available cell functions and topology. Placement changes the physical context in which that topology is measured. Critical-path analysis therefore needs both the mapped netlist’s structural dependencies and the timing constraints under which the path is reported. A shared cell sequence alone is insufficient to explain its delay.

Multi-corner static timing analysis

Setup and hold are evaluated separately across declared timing scenarios. Corner identity, constraints and analysis fidelity must remain visible. Worst negative slack, total negative slack and violating-endpoint counts describe different aspects of timing; none can substitute for a complete scenario matrix.

Reproducible comparisons

Comparisons require identifiable parent and candidate checkpoints, common inputs and constraints, known tool configurations and consistent measurement fidelity. A selected candidate must be reloaded and remeasured independently of its search record. Timing and area trade-offs remain distinct from the question of whether the implementation satisfies every declared constraint.

Functional preservation versus physical feasibility

A legal placement with improved measured timing does not prove preservation of sequential behaviour. Structural resynthesis requires a separate functional-equivalence argument with established assumptions. Physical measurements and successful software tests cannot supply that proof. The available research does not establish functionally verified structural resynthesis.

Bounded example: GCD / Nangate45

Objective. Examine whether sampled timing paths expose shared mapped structure that could guide a subsequent physical-design investigation.

Setup. The recorded analysis uses a GCD implementation in an OpenROAD flow with the Nangate45 standard-cell library and typical, fast and slow technology corners. Its timing evidence is based on legalised placement with estimated placement parasitics. It is not extracted post-route or signoff-quality verification.

Evidence scope. The following counts are reported in Viridian’s engineering records and were checked against the recorded physical-constraint analysis. They describe one implementation record, not a new run or a cross-design benchmark. No EDA experiments were executed to produce this page.

Recorded analysis scope — one GCD / Nangate45 implementation

Recorded quantityCount
Standard-cell instances469
Pins1,149
Design ports54
Declared nets426
Complete aggregate connectivity records417
Timing-scenario records3
Sampled timing paths60
Slow-corner setup-violating endpoints53
Sampled worst setup paths with a common segment10
Cells in that shared segment8

Interpretation and maximum conclusion

Ten sampled worst slow-corner setup paths share an eight-cell structural segment. Repeated participation in these paths makes that segment a plausible region for investigation. It does not show that the segment causes the timing failures, nor that modifying it would improve the implementation.

For this recorded GCD / Nangate45 implementation, the ten sampled worst slow-corner setup paths contain a common eight-cell segment. The observation supports targeted investigation of that region, subject to incomplete connectivity and timing evidence. It establishes no causal bottleneck or optimisation improvement.

The ten sampled paths account for only ten of the 53 violating endpoints. The result cannot explain the remaining 43 endpoints or be treated as exhaustive critical-path coverage.

Limitations and scientific boundaries

Nine of the 426 declared nets lack complete aggregate connectivity records. Full port-to-net correspondence was not established for the 54 design ports. These gaps restrict complete boundary and connectivity analysis; missing relations are not inferred from names.

The saved path evidence lacks point-level arrival, required time, slew and transition data, together with a cell-versus-net delay partition. Incomplete path-level timing information limits causal interpretation of the shared segment. Three scenario records and 60 sampled paths do not establish exhaustive timing coverage.

The available evidence does not establish complete timing closure, routed or signoff-quality verification, power improvement, functionally verified structural resynthesis, or generalisation to other designs. It provides no basis for claiming superiority over commercial EDA tools, verified commercial performance improvements or production-ready semiconductor optimisation.

OpenROAD and Nangate45 identify the experimental tooling and library. Their use implies no endorsement or partnership. This page is a bounded research summary, not a peer-reviewed publication or a production qualification certificate.

Evidence needed for a stronger conclusion

A causal investigation would need more complete connectivity and port binding, richer path-level timing data, and a controlled intervention evaluated from a common parent checkpoint. The intervention would require comparable setup and hold analysis across every declared scenario, area and placement checks, and separate functional-preservation evidence wherever logic is changed.

A fair optimisation comparison would additionally need a frozen protocol, matched budgets and retained failures. Generalisation requires other designs and independently comparable execution. These are requirements for stronger claims, not results already established by this analysis.