part1_fanout_escape module#
JS4 Part 1: fanout and escape, the layout’s first step, in code.
JS4FanoutCircuit takes the connected board of base
and lays down one via and routing strategy for both parts:
Placement. The FPGA stays where the base design places it. The memory sits below the FPGA’s LPDDR4 banks at (6.4509, -17.3772) mm on the board, turned 180 degrees, a design choice: to balance length across the data lanes (
MEMORY_AT).Layers, from UG863’s layout rules as
constraintlists them (a byte’s data and strobe on one internal layer, command, address and clock on one internal layer, data no deeper than 45 mil) and DIFF-85’s inner layers, L3, L5 and L7. Channel B’s bytes take L3_SIG1 (CHANNEL_B_LAYER) under the FPGA and between the parts, so channel A’s go to L5_SIG2 (14.2 mil below the top of TOP’s copper, by the stackup table), reached by the fanout’s through vias: the substrate has no TOP to L5 laser via. Both channels’ CA, CS, CKE and CK go on L7_SIG3.Fanout. One through via at the centre of the pad of every FPGA ball but those the pinout names NC, and of every memory ball but NC and DNU. At the FPGA that includes the balls no net reaches yet (unused I/O and transceiver pins, the system clock, VCCINT_SENSE), a design choice that keeps every pin available to a later part. The NC balls get none: the kit leaves them unconnected, so no later part needs them (
FanoutFpga.fpga_balls()).Ground. A GND pour over the board outline on each of the eight ground layers, L2, L4, L6, L8, L11, L13, L15 and L17, held by the ground vias.
Escape. Channel B’s two DQS pairs, the ones Part 4 will route, leave each part’s ball field on L3 on route segments tagged
BallFieldPairTag, from the pair’s two fanout vias through aPairInsertionbeside them to aPairPointjust outside the field, where routing continues on the full pair structure.ball_field_rules()gives the tagged segments the ball-field pair structure. Those two are the escape’s only control points: the insertion where its two legs couple, and the pair point where the tag ends. No other pair escapes in code.Decoupling. The base design’s capacitors stay unplaced, and the FPGA’s circuit never creates the 27 that Part 2 (
part2_power) would have nowhere to put where they decouple anything (FanoutFpga.decoupling_layouts()). Part 2 pours five rails, and puts a capacitor of 10 uF and up in a ring outside the ball field on its own via to its rail’s pour, which on any other rail would join nothing. So on each FPGA rail Part 2 does not pour (POURED_RAILS), the design has no capacitor of 10 uF and up (19 of them), and none of the eight below 10 uF that Part 2’s field placement finds no free pair of ball vias for (NO_FIELD_PAIR), which would go to the ring too. It holds 59 of the base design’s 86 capacitors: 44 of the FPGA’s 71 and all 15 of the memory’s, whose rails Part 2 pours.
Each part’s copper sits in the part’s own circuit, so it moves when the
part’s circuit moves: the FPGA’s vias, insertions, pair points and escape
routes in FanoutFpga, the memory’s in FanoutMemory.
The memory sits at its circuit’s origin turned 180 degrees, and its circuit
is what is placed. The pours cover the whole board and stay in
JS4FanoutCircuit.
Escape positions. Each escape’s insertion and pair point sit at the
position and heading reference_escapes gives, in the circuit of
the part they belong to: the control points of a reference layout of
channel B’s two byte lanes on this board, memory at (6.4509, -17.3772)
turned 180. The FPGA’s are keyed by byte slot
(FanoutFpga.escape_pairs()), the memory’s by byte lane
(FanoutMemory.escape_pairs()).
Pair polarity, by jitx.controlpoint.PairInsertion and
jitx.controlpoint.PairPoint: an insertion’s coupled end joins
the back of a pair point with the same invert flag, and two pair points
whose flags differ join front to front. Each part’s insertions and pair
points share one flag, inverted at the FPGA and not at the memory
(FPGA_INVERT), which at the saved headings keeps each
insertion’s P leg on its P via’s side. Whichever of channel B’s bytes the
pin assignment puts in either of its slots, its two pair points then join
without P and N crossing between the parts.
Via size. Each fanout via is THVia_6p0, the smallest drill the
substrate offers TOP to BOTTOM, a design choice: it leaves the most room
between pads for the ball-field routes. Its pad is 0.252 mm (the drill plus
two IPC-2221B 0.050 mm rings,
via_pad()), inside the
VP1002’s 0.51 mm land and the memory’s 0.40 mm ball pad. With a full pad on
both neighbours, the narrowest case on an inner layer, the gaps between
neighbouring pads and what fits in them are:
FPGA, 0.92 mm pitch both ways: 26.28 mil. The ball-field pair fits (17.2 mil,
DRS_85_FIELD), as do twoSE_46lines (24.8 mil: 3.1 mil wide, 6.2 mil clearance).Memory columns, 0.8 mm pitch: 21.56 mil. The ball-field pair fits, as does one
SE_46line (15.5 mil), but not twoSE_46lines.Memory rows, 0.65 mm pitch: 15.65 mil. One
SE_46line fits; the pair does not.
Two more design choices at the memory. Its escapes take DRS_85_FIELD,
documented for a 0.92 mm field, at its 0.8 by 0.65 mm field, because the
pair fits between its columns and the substrate has no structure for that
pitch. And they take UG863’s under-device spacing, because the
BallFieldTag that gives them the narrower
structure is also the spacing rules’ under-device tag.
The design applies no SI constraints; part3_si_constraints adds
them on top of Part 2’s circuit (part2_power), which builds on
this one.
- THROUGH_VIA#
The via in every fanout pad: the smallest plated drill on the substrate’s TOP to BOTTOM chart, 6.0 mil, 0.252 mm pad. Smallest is a design choice: it leaves the most room between pads for the ball-field routes.
- CHANNEL_B_LAYER = 2#
L3_SIG1, a design choice, the layer the reference layout of channel B’s byte lanes puts them on.
- Type:
Channel B’s bytes
- GROUND_LAYERS = (1, 3, 5, 7, 10, 12, 14, 16)#
L2_GND, L4_GND, L6_GND, L8_GND_1OZ, L11_GND_1OZ, L13_GND, L15_GND and L17_GND: every ground layer of the substrate’s stackup.
- MEMORY_AT = (6.4509, -17.3772)#
Where the memory’s circuit sits on the board, mm, at 0 degrees, so the memory, turned 180 degrees in it, sits there turned 180: below the FPGA’s LPDDR4 banks with channel B’s end toward them. A design choice, to balance length across the data lanes.
- POURED_RAILS = frozenset({'VCCAUX', 'VCCINT', 'VCCO_700', 'VCCO_701', 'VCCO_702', 'VCC_IO', 'VCC_SOC'})#
The FPGA rails Part 2 pours, by the set names of
fpga_sets(): the VCCO of XPIO banks 700 to 702 (on the 1.1 V net), VCCINT, VCC_SOC with VCC_IO (one net) and VCCAUX. On every other FPGA rail the design creates no capacitor that Part 2 would put in its ring (module docstring, “Decoupling”).
- NO_FIELD_PAIR = mappingproxy({'VCCO_503': frozenset({1}), 'VCC_PMC': frozenset({1}), 'VCC_PSLP': frozenset({1}), 'VCCAUX_SMON': frozenset({1}), 'GTY_AVCC_RS': frozenset({2}), 'GTM_AVCC_RN': frozenset({2}), 'VCCINT_GT_R': frozenset({1, 2})})#
The capacitors below 10 uF on rails Part 2 does not pour that its field placement finds no free pair of ball vias for, by rail name and place in the rail’s set: the 4.7 uF of VCCO_503, VCC_PMC and VCC_PSLP; VCCAUX_SMON’s 100 nF; the 470 nF of GTY_AVCC_RS and GTM_AVCC_RN; and VCCINT_GT_R’s 4.7 uF and 470 nF. Each belongs to a rail of one to three balls whose few rail and ground pairs are taken. Whether a capacitor finds a pair is known only once Part 2 places the field, so the result is worked out once and written here as data; the repository’s Part 2 tests place every set whole and check it.
- FPGA_INVERT = True#
The
invertflag of every FPGA pair point and insertion. Every memory pair point and insertion takes the opposite, so the two pair points of a pair join front to front (module docstring, “Pair polarity”).
- pad_centres(component)[source]#
Position of the centre of each port’s pad on a placed
component, in the frame of the circuit that holds it, read in one pass over its pads (the first pad, for a port with several).
- via_in_pad(centre, ball, kind)[source]#
A
kindvia at the centre of the ball’s pad, attached to the ball. The caller stores both in the circuit that holds the ball’s part.
- attached_via(attachments, ball)[source]#
The via attached to
ballamongattachments.- Return type:
- Parameters:
attachments (list[PortAttachment])
ball (Port)
- tagged_pair_routes(p_via, n_via, insertion, point, layer)[source]#
The three segments of a pair’s escape, each tagged
BallFieldPairTag: its two vias’ legs into the insertion and the coupled segment from the insertion to the pair point.
- pair_escape(p_via, n_via, saved, layer, invert)[source]#
A pair’s insertion and pair point where
savedplaces them, both withinvert, and its three tagged segments. The pair point shares the insertion’s flag, so the insertion’s coupled end joins its back.- Return type:
- Parameters:
p_via (Via)
n_via (Via)
saved (PairEscape)
layer (int)
invert (bool)
- class FanoutFpga[source]#
Bases:
JS4FpgaThe FPGA’s circuit with a via in every ball’s pad but the NC balls’, the pair escapes out of its ball field, and none of the capacitors Part 2 would have nowhere to put (
decoupling_layouts()).- fanout_attachments: list[PortAttachment]#
- escape_insertions: list[PairInsertion]#
- decoupling_layouts()[source]#
Each FPGA set’s layout, read before any capacitor is created: on a rail outside
POURED_RAILS, leave out every capacitor of 10 uF and up and those ofNO_FIELD_PAIR, since in Part 2’s ring their rail via would join nothing (module docstring, “Decoupling”). Every other capacitor is created, each below 10 uF tied to its ball by a ShortTrace.- Return type:
- fpga_balls()[source]#
The FPGA’s fanout balls: every ball but those the pinout names NC, whether a net reaches it yet or not, a design choice that keeps every pin available to a later part: a later part may assign an unused I/O ball. UG863 v1.12 (“Unused VCCO Banks and NC Pins”): “Pins that are labeled as “NC” in the package file can either be left unconnected or tied to ground.” The kit leaves them unconnected, a design choice, so no later part needs them.
- escape_pairs()[source]#
Each escaped pair’s byte slot and its (P, N) FPGA balls: the DQS of channel B’s two byte slots, in slot order.
- escape(p, n, saved)[source]#
Escape one pair out of the FPGA’s ball field: its two vias to a
PairInsertioncoupled to aPairPoint, both wheresavedplaces them, each segment onCHANNEL_B_LAYERtaggedBallFieldPairTag, all withFPGA_INVERT.- Return type:
- Parameters:
p (Port)
n (Port)
saved (PairEscape)
- class FanoutMemory[source]#
Bases:
JS4MemoryThe memory’s circuit with the memory at its origin, turned 180 degrees, a via in the pad of every ball but NC and DNU, and channel B’s two DQS pairs escaped out of its ball field.
- fanout_attachments: list[PortAttachment]#
- escape_insertions: list[PairInsertion]#
- escape_pairs()[source]#
The memory’s (P, N) balls of each escaped pair: DQS_t and DQS_c of channel B’s byte lanes 0 and 1 (
ch[1]).
- escape(p, n, saved)[source]#
Escape one pair out of the memory’s ball field toward the FPGA: its two vias to a
PairInsertioncoupled to aPairPoint, both wheresavedplaces them, each segment onCHANNEL_B_LAYERtaggedBallFieldPairTag, all with the opposite ofFPGA_INVERT.- Return type:
- Parameters:
p (Port)
n (Port)
saved (PairEscape)
- class JS4FanoutCircuit[source]#
Bases:
JS4BaseCircuitJS4BaseCircuitwith Part 1’s placement, fanout, ground pours and channel B’s two DQS escapes.
- lane_balls(lane)[source]#
Every ball of a memory byte lane: the eight DQ, DQS_t, DQS_c and DMI.
- Return type:
- Parameters:
lane (LPDDR4Lane)