Process files and extraction

EMIR keeps every process number in one YAML file per PDK, the process file (--process); no limit lives in the code. This page says where the files come from, how to write one (from the PDK's files, an ICT, or typed by hand from the design manual), every key, what the extracted DSPF must contain, and how magic extracts one on SKY130.

Screening, not sign-off. A process file is a reading of the PDK's data, not the foundry's qualified rule deck.

A minimal file

pdk: my_pdk (from the tech LEF, rev X)
em_tref: 105.0                  # C, where the LEF quotes its limits
temp_reference: 25.0
layers:
  M1:
    gds: [10, 0]
    sheet_res: 0.08             # ohm/sq at 25 C
    tc1: 0.0035
    min_width: 0.1
    jmax_avg: 1.5               # mA/um at 105 C
  M2:
    gds: [12, 0]
    sheet_res: 0.08
    tc1: 0.0035
    min_width: 0.1
    jmax_avg: 1.5
vias:
  V1:
    gds: [11, 0]
    below: M1
    above: M2
    res: 4.0                    # ohm per cut
    vcwidth: 0.1                # um
    imax_avg: 0.2               # mA per cut

(The numbers are invented; a real file cites a source for each.) Units: lengths in um, resistance in ohm (per square for wires, per cut for vias), limits in mA per um of drawn width for wires and mA per cut for vias, temperatures in C. emir schema process prints the JSON Schema; pdks/TEMPLATE.yaml has every key with comments.

The files in pdks/

file	PDK	metal stack	limits quoted at
sky130.yaml	SkyWater SKY130 (sky130A)	li1, met1 to met5; mcon, via, via2 to via4; and the magic: section	90 C
gf180mcu.yaml	GlobalFoundries GF180MCU (gf180mcuD)	Metal1 to Metal5; Via1 to Via4	125 C
ihp_sg13g2.yaml	IHP SG13G2	Metal1 to Metal5, TopMetal1 and 2; their vias	105 C

They are built only from the PDKs' own public files (Apache-2.0, the open PDKs' licences), and every value names its source as [FILE]:line. With the PDKs installed, pytest tests/test_pdks.py -k citations checks every cited line still holds its value. pdks/README.md lists what each PDK does not publish (no peak limits on any of the three, no rms limits on vias): those metrics are reported as not checked.

Writing one for your PDK

1. Copy pdks/TEMPLATE.yaml and change pdk: (a name containing "illustrative" warns that the limits are not real, EMIR-W306).
2. Fill each conductor layer from the PDK's documents: sheet resistance at temp_reference, the TCRs, and the EM limits in mA/um at the temperature in em_tref. Fill each via: below, above, ohm per cut, the cut side vcwidth, the limits in mA per cut. Cite each value in a comment.
3. Leave out what the PDK does not publish: a missing limit is reported as not checked. Never estimate one.
4. Map the extractor's layer names that do not match with dspf_layers (m1_conn_a: Metal1, or 1: Metal1 for a $lvl=1 level number).
5. Check it: emir check --process my_pdk.yaml alone, then emir check <dspf> --process my_pdk.yaml, and one hand-checkable wire at its limit (sky130's met1 allows 2.8 mA/um: a 0.14 um wire may carry 0.392 mA; tests/test_pdks.py runs that strap).

From a foundry ICT: emir import-ict --ict <pdk>.ict --emir my_pdk.yaml -o new.yaml fills the EM limits, TCRs, jmax_factor and em_tref onto the geometry already in the file. A J x (w - dw) limit becomes J with width_loss_um: dw; the long-wire branch is taken (the conservative one); a via rms or peak limit the ICT does not give is not written. A file made from a licensed PDK's data stays with that PDK installation, never in this repository (the leak guard refuses ICT, tech-file and layer-map files).

Unit traps: a density per cross-section (MA/cm^2) is not mA/um (multiply by the metal thickness: 1 MA/cm^2 = 10 mA/um^2); a via limit or resistance per array is not per cut; vcwidth is the cut's side, not its area; em_tref (where the limits hold) is not temp_reference (where the sheet resistance holds); a LEF's fixed rms limit is jmax_rms, an ICT's self-heating constant jmax_rms_k, never both (EMIR-E310); give the unidirectional DC average limit, not a bidirectional one. emir check flags a value outside its quantity's plausible range (EMIR-W321: an A-for-mA or nm-for-um slip), but not a factor of ten.

Entering EM data by hand

When no readable process data exists (not in pdks/, and no ICT, or an encrypted one: emir import-ict refuses it, EMIR-E325, and never tries to decrypt it), the person types the EM data from the design manual:

emir process new --from-dspf block.dspf -o my_pdk.yaml     # the fill-in file
emir process check my_pdk.yaml --dspf block.dspf            # what is still missing

• The fill-in file covers exactly the layers and vias the DSPF uses, every value null with its unit and a sources: slot, the via stacking it inferred marked to be checked, and provenance: hand-entered. It never overwrites a file (EMIR-E326).
• Type each value as the manual prints it, with its source (sources: {jmax_avg: "DM rev 2.1, table 9-4, p. 112"}); leave null what the manual does not give. At a terminal, emir process new ... --interactive asks for each value; an agent never drives it (EMIR-E327).
• The check lists what is still empty (EMIR-E320 for a required value), without a source (EMIR-W320), unconfirmed stacking (EMIR-W324), likely unit slips (EMIR-W321) and limits out of order (EMIR-W322). A likely slip on a temperature, the derating or a limit blocks the run (EMIR-E331) until it is corrected or the person lists its path under unit_confirmed:. It must pass before any run.
• Every report says per layer that the limits were typed by hand, and from which document (limits_provenance).

Two forms the manuals print are read as printed. A width loss, I_max = J x (W - dW), is width_loss_um: dW (for J = 2.0 mA/um and dW = 0.02 um, a 0.30 um wire may carry 0.56 mA, 7 % stricter than 0.60 mA; a wire no wider than dW is refused, EMIR-E323). Black's equation, MTTF = A J^-n exp(Ea / (k T)), is derating: {model: black, ea_eV: Ea, n: n}: the avg limits scale by exp[(Ea / (n k)) (1/T - 1/Tref)] (Ea = 0.9 eV, n = 2, from 105 C to 125 C: x0.4997); below em_tref they are held as quoted (EMIR-W326). emir validate holds both to a closed form.

Top level

key	required	meaning
pdk	no	the name reports show; its first word is the PDK's directory under $PDK_ROOT (magic's startup file)
provenance	no	hand-entered, imported-ict, public-pdk-data or mixed: reports say it per layer; a hand-entered value with no source gets EMIR-W320
sources	no	where em_tref, temp_reference, jmax_factor and derating came from
em_tref	yes	C, the temperature the average limits hold at, and the default analysis temperature (no default: EMIR-E312)
temp_reference	no	C, where sheet_res holds (default 25)
warn_fraction	no	the ratio from which an item is WARN (default 0.8)
jmax_factor	no	{C: factor}, the derating of the average limits, 1.0 at em_tref
derating	no	{model: black, ea_eV, n}, Black's equation; give it or jmax_factor, never both (EMIR-E321)
unit_confirmed	no	paths of values outside their usual range that the person checked against the manual
dspf_layers	no	{token or $lvl number: name}, for extractor names that do not match
magic	no	the magic route (below)
layers	yes	the conductor layers (at least one)
vias	no	the cut layers
ict_conductor_pattern	no	emir import-ict: the ICT block name for a layer (default {layer}_conn)
ict_process	no	written by emir import-ict, for the record

An unknown key inside a layer or via is a FAIL (EMIR-E316: a misspelt limit would go unchecked); an unknown top-level key warns (EMIR-W314). Any value may be null: a limit left empty is not checked, a required value left empty is a FAIL (EMIR-E320).

A conductor layer

key	unit	meaning
gds	[layer, datatype]	for the person; no path reads it
sheet_res	ohm/square	at temp_reference (required)
tc1	1/C	the linear TCR: R(T) = R (1 + tc1 dT + tc2 dT^2), dT from temp_reference
tc2	1/C^2	the quadratic TCR
min_width	um	the PDK's rule, for the record
jmax_avg	mA/um	the DC average limit at em_tref, derated at other temperatures
jmax_rms	mA/um	a fixed rms limit, as a tech LEF gives it
jmax_rms_k	(mA/um)^2 per C	an ICT's self-heating form: the limit is sqrt(jmax_rms_k x delta_T)
jmax_peak	mA/um	the peak limit
width_loss_um	um	the dW of I_max = J x (W - dW)
sources		where each value came from
provenance		this entry's own, in a mixed file
estimate		true flags the limits as estimates in every report
ict_name		emir import-ict: this layer's ICT block

A via

key	unit	meaning
gds	[layer, datatype]	the cut layer
below		the conductor under the cut (required)
above		the conductor over the cut (required)
res	ohm per cut	required; not scaled with temperature
vcwidth	um	the cut's side: turns an extracted cut area ($A) into a count. Without it such a via is checked as one cut (EMIR-W308)
imax_avg	mA per cut	the DC average limit at em_tref, derated at other temperatures
imax_rms	mA per cut	the rms limit
imax_peak	mA per cut	the peak limit
estimate		flags the average limit as an estimate
rms_peak_estimate		flags imax_rms and imax_peak as estimates: they are not applied
sources		as for a layer
provenance		as for a layer
ict_name		emir import-ict: the ICT via block

The magic section

key	meaning
binary	the magic to try, a magic program's bare name on $PATH (a path is refused, EMIR-E648)
layers	magic's layer names onto the file's (locali: li1, viali: mcon on sky130)
devices	each device model's terminal names in magic's order ("sky130_fd_pr__nfet_*": [d, g, s, b]); others get t1, t2 ... (EMIR-W643)
sources	where the values came from

--process FILE wins; then $EMIR_PROCESS; then the site file's process: ($EMIR_SITE, else ~/.emir/site.yaml, never committed).

The DSPF

EM needs each wire's width (J = I / W) and each via's cut count: SPEF has no geometry, a DSPF with geometry annotation does. emir check <dspf> --process <file> names every gap before anything is computed.

record	what EMIR takes from it
*|OPERATING_TEMPERATURE	the extraction temperature (resistance is scaled to the analysis temperature)
*|NET, *|P	the net, and the port where it is fed (the IR reference)
*|I (node inst pin dir cap x y)	where a device pin lands: simulated currents attach here (without x y the record is dropped)
*|S (node x y)	sub-node coordinates, for the maps
R... n1 n2 value $layer $W= $L= ...	a wire: its layer and drawn width ($W or $SI_W); without a width, a FAIL (EMIR-E210)
R... $cont $N= / $VC= / $A=	a via: its cut count, or its cut area turned into a count with vcwidth

Extraction settings, whatever the extractor: no series reduction (one resistor per drawn segment, so a neck is not averaged away), width and layer on every wire, a cut count or area on every via, node and instance-pin coordinates, the extraction temperature stated, ideally the EM analysis temperature. Quantus: Cadence; Calibre and StarRC: notes in agent/cadence/ (UNVERIFIED). An R-only netlist is not the netlist to simulate: it has no capacitance.

currents	the DSPF EMIR screens	the netlist you simulate
a simulation (rcc)	the RC extraction	the same RC extraction, every resistor current saved
DC (--currents)	R-only or RC	none

SKY130: magic extracts the DSPF

On SKY130, with a GDS and no DSPF, magic extracts one: emir extract --tool magic has magic keep every resistor as one drawn segment, with its layer, width, length and place, and writes the DSPF.

emir next --gds blk.gds --process pdks/sky130.yaml --cell blk --format json
emir extract --tool magic --gds blk.gds --cell blk --process pdks/sky130.yaml --out emir_out/extract_blk
emir next --dspf emir_out/extract_blk/blk.dspf --process pdks/sky130.yaml --netlist tb.cir \
    --gds blk.gds --cell blk --format json

<out>/<cell>.dspf is the network, and the RC netlist an rcc testbench includes. From there the flow is the one on Simulators. With --html, emir run --gds blk.gds --cell blk draws the layout under the chart.

The magic it needs. A stock magic writes each resistor's value only. The geoout patch adds extresist geoout on, which writes each resistor's layer, direction, width, length and end points (<cell>.res.geo). It is public: branch extresist-geoout of codervijay/magic, offered upstream as RTimothyEdwards/magic#562 (open).

git clone -b extresist-geoout https://github.com/codervijay/magic magic-geo
cd magic-geo && ./configure --prefix=$HOME/magic-geo && make && make install
export EMIR_MAGIC=$HOME/magic-geo/bin/magic

EMIR tries --magic, $EMIR_MAGIC, the process file's magic: binary:, the site file's magic:, then magic-8.3.515-geo and magic on $PATH, and uses the first that answers extresist geoout; with none it exits 3 (EMIR-E643) and says how to build one. The examples were made with magic 8.3.515 and the patch (the branch is on 8.3.684; the test cells in tests/data/magic_tiny give the same resistor network with both). magic's technology is $PDK_ROOT/<pdk>/libs.tech/magic/<pdk>.magicrc (--magicrc overrides it).

What it writes. The cell is flattened first (a hierarchical layout extracts as the flat one would). Every conductor magic extracts is in the network (contacts, local interconnect, diffusion, poly, wells); a layer with no limit carries its current in the IR drop and is reported as not checked (on sky130: li1 and below). Each device terminal is a *|I record named from magic: devices:. A port drawn with several pins gets a node per pin, tied to the port by 0 V sources (--port-pins tied, the default), so every pin is fed. Capacitors are counted once. The DSPF writer and the geometry reader are ported from the Skycadeda Analog Framework (SAF), with credit and the licence in THIRD_PARTY_NOTICES.md.

Example 08. examples/08_open_layout runs the route on a synthetic SKY130 cell: four two-finger NMOS whose out and vss each reach their pad through one via2 cut. magic extracted 1135 resistors; the gate is pulsed, Xyce 7.10, 90 C:

run	out: IR, worst ratio	vss: IR, worst ratio
rcc	6.6910 mV, 1.3566x (via2: 1 -> 2 cuts)	4.8448 mV, 1.3566x

rcc's 440 checked branches equal the simulator's own currents to 7.8e-13, and its IR the simulator's own node voltages to 3.2e-7.

Example 08: the chart of its rcc run (synthetic layout, the SKY130 limits of pdks/sky130.yaml).

Limits of the route: it needs the patched magic until the patch is in a magic release; tested on sky130 only (gf180mcu and IHP SG13G2 not tried), on small cells; a wire on a diagonal shape is checked at magic's tile extent and listed as not checked (EMIR-W647); a net magic leaves whole has nothing to check (EMIR-W645); label every net.