Getting started

Install EMIR, prove the install, and screen the bundled sample: a minute or two, offline after the install.

Screening, not sign-off. The sample's limits are invented and its verdict is FAIL on purpose. EMIR is a research project, provided as is, with no warranty or guarantee of accuracy: test and verify its results for your process, tools and design before relying on them.

Install

Linux (x86_64 is what is built and tested) and Python 3.10 or newer. The core needs only numpy, SciPy and PyYAML, and no licence.

From a clone (recommended, and what an agent uses):

git clone https://github.com/skycadeda/emir emir
cd emir
python3 -m venv .venv
. .venv/bin/activate
pip install -e .              # the core; pip install -e '.[all]' adds every extra

A clone also runs without installing: bin/emir uses the first python3 (3.10 or newer) that has numpy, SciPy and PyYAML, or $EMIR_PYTHON.

The offline bundle, for a design server with no network: each GitHub Release carries emir-<version>-linux-x86_64.tar.gz and SHA256SUMS (a standalone Python with the packages, bin/emir, the docs, agent/, pdks/ and examples/). Check it with sha256sum -c SHA256SUMS --ignore-missing, unpack it and run emir-<version>/bin/emir selftest. EMIR_ALLOW_LOCAL_BUNDLE=1 sh scripts/build_bundle.sh builds one yourself.

The wheel: pip install emir-<version>-py3-none-any.whl (from a Release, or python -m build in a clone). EMIR is not on PyPI. A wheel holds the package and the bundled sample, not agent/, examples/ or pdks/: advice that cites one of those files names its page in the repository. emir info prints where the sample is (sample_dir).

extra	adds	for
report	matplotlib	the EM/IR chart of --html
klayout	the klayout package	the KLayout marker database, the layout under a chart, the top cell for emir extract --tool magic
mcp	the MCP Python SDK (1.21.1 or newer)	emir mcp
all	all of the above	everything a user runs
dev	the above plus gdstk, jsonschema, pytest, ruff, pre-commit, build	the tests
docs	MkDocs with the Material theme	this site

External tools are needed only by their flows: Xyce or ngspice, HSPICE, Spectre, PVS and Quantus, magic with the public geoout patch, KLayout. emir info says what it found.

Prove the install

emir selftest          # 19 checks on the bundled sample, offline, seconds: SELFTEST PASSED
emir validate          # the 6 analytic validation cases
emir info              # versions, and where each external tool was found

Run emir selftest before you touch real data, and after any upgrade. Before the first run that uses a simulator's results on a machine, verify its route there: emir verify-sim --sim xyce (your simulator's name; see Simulators).

The first run

The bundled sample is a 60 um Metal1 rail, 0.75 um wide, in 12 segments of 1 ohm, with four 1 mA loads tapping it through contacts, and invented round-number limits. From the root of the clone:

emir run emir/data/sample/rail.dspf --process emir/data/sample/process.yaml \
    --currents '{"VDD": 0.004, "MP0:s": -0.001, "MP1:s": -0.001, "MP2:s": -0.001, "MP3:s": -0.001}' \
    --summary-out emir_out/summary.json --html emir_out/html
emir report --from emir_out/summary.json

--currents gives DC currents in amps flowing into the net: 4 mA in at the port VDD, 1 mA out at each load's pin. The input check warns (EMIR-W306) that the sample's limits are not real; then the report gives each failure with its fix:

| VDD | Rm1 | (2.500, 0.000) | Metal1 | wire | avg | 1.667x | Metal1: widen 0.750 -> 1.255 um (x1.67) |
| VDD | Rc3 | (60.000, 0.000) | Cont | via | avg | 1.333x | Cont: 1 -> 2 cuts (+1) |
- VDD: 54.000 mV (from the feed)

By hand: the first three segments carry all 4 mA in 0.75 um, 5.33 mA/um against the sample's 3.2 mA/um: 1.667x. The last contact carries 1 mA through one cut against 0.75 mA per cut: 1.333x. The IR drop is 4 + 3 + 2 + 1 mA through 3 ohm each (30 mV), plus 1 mA through the last contact's 24 ohm (24 mV): 54 mV. Open emir_out/html/index.html for the map of the net.

Next

emir next reads what you have and prints the commands to run, in order, each with its reason:

emir next --dspf block.dspf --process pdk.yaml --cc-sim result.raw --period <T>

With a simulation result it plans rcc: the check and the run on every branch current the result holds, or the save lines and the simulation when it lacks them. A transient needs the block's period (--period): without it, the plan asks for it.

you have	read
a DSPF and a simulation result	Method and accuracy
Xyce, ngspice, Spectre or HSPICE	Simulators
a PDK with no process file yet, or only a GDS on SKY130	Process files and extraction
Cadence tools	Cadence
an agent to drive all this	AGENTS.md

The worked examples in examples/ run offline with bash examples/<name>/run.sh and compare their results with golden files.