OpenCell-5

A 5nm-class statistical SDK — an open, IP-free standard-cell deck that gives you a representative feel of a 5nm design flow when you don't have access to a real foundry PDK. It is derived by statistical scaling of the silicon-validated SkyWater sky130 PDK (Apache-2.0), and it runs the full open OpenROAD/ORFS flow: synth → floorplan → place → CTS.

→ Just want to run it? See docs/QUICKSTART.md. The deck ships committed; a user never touches sky130.

docker pull openroad/orfs:latest
flow/run_orfs.sh opencell5 gcd cts          # run a design to the statistical endpoint
flow/ppa.py opencell5 gcd                    # read area / fmax / power

OpenCell-5 rewrites every numeric field of the sky130 cell library — delays, transitions, setup/hold, leakage, power, cap, area — and the physical LEF geometry to a 5nm-class operating point, then drives real designs through OpenROAD to a post-CTS, router-free statistical endpoint (a real clock tree + RC-estimated timing). That endpoint is where the honest PPA lives; both this deck and its 7nm sibling reach it cleanly.

The 5nm-class scale factors (see scaling/scale_factors_5nm.json, citations in docs/SOURCES.md):

• Not sign-off. The endpoint is post-CTS; no detailed route, no DRC/LVS-clean GDS. Manufacturability is not the goal — a representative statistical feel is.
• Not foundry-correlated. Statistically consistent with published 5nm scaling; not equivalent to any specific TSMC/Samsung/Intel node.
• Not silicon-validated at 5nm. Silicon-validated at 130nm (sky130); the structural correctness inherits, the numeric magnitudes are projected.
• Not GAA. sky130-derived FinFET-class scaling; real 5nm is still FinFET, but 3nm/2nm gate-all-around is out of scope.

├── README.md
├── docs/QUICKSTART.md            run it yourself (no agent, no sky130)
├── docs/RUN_ORFS.md              the containerized ORFS wrapper + gotchas
├── docs/METHODOLOGY.md           scaling theory, derate, corner derivations
├── flow/
│   ├── run_orfs.sh               drive <platform> <design> to a stage (Docker)
│   ├── ppa.py                    single-platform PPA readout
│   ├── statppa.py                correlate any two platforms (--platforms A B)
│   └── tighten.py                push a design's clock to its true fmax floor
├── scaling/
│   ├── scale_factors_5nm.json    the opencell-5 definition (5nm factors)
│   ├── build_opencell5.sh        rebuild the deck from source (maintainers)
│   ├── scale_lib.py              sky130 .lib -> node-class .lib
│   └── gridlock_lef.py           site/DB-grid-locked LEF scaler
├── platforms/opencell5/          the 5nm-class deck (committed)
└── designs/{opencell5,src}/      design configs + RTL

Apache-2.0. Upstream sky130 Apache-2.0 attribution is preserved verbatim in every .lib the scaler produces. See LICENSE.

github.com/rtoley/Opencell-5