Repository navigation
Conversation
The ASAP7 PDN scripts place power vias and stripes where the design rule
manual's grid, track and via-width rules cannot hold:
- pdngen stacks the M2-to-M5 connections on the M2 rails, every 270 nm,
which is off M4's 24 nm grid; the stacked M3/M4 pads are wider than
their cuts. Now -min_width_layers {M3 M4} -ongrid {M4}.
- 0.12 um M5 straps take V4 cuts narrower than the strap. The straps are
now minimum width (0.024 um) and -snap_to_grid, at a fifth of the pitch
(1.08 um, 0.432 um at the BLOCKS top level) so the M5 metal is the same.
M6 straps are -snap_to_grid too.
- Rings: the new ring_offsets.tcl computes per-side core offsets that put
every ring wire on its layer's tracks. The BLOCKS ring's M5 is 0.408
(0.504 is over M5's 0.48 maximum) and its M6 0.288, the width
M6_M5widePWR1p152 makes a flush V5 for; BLOCK's ring M5 is minimum width
for the V4s at its corners.
Checked with gdscheck's asap7 openroad suite (the DRM, less the rules the
tech LEF does not give the router): gcd 602 -> 0, aes 14449 -> 0, aes-block
114821 -> 15756 (what remains is macro placement off the fin/gate grid and
tech LEF via and width gaps). Detailed routing has no DRC errors on all
three. Worst-case static IR drop (VDD) is lower than before:
gcd 0.61% -> 0.40%
aes 4.56% -> 2.80%
aes-block 0.44% -> 0.17%
gcd and aes built with SYNTH_USE_SYN=0; syn currently writes DEF at 100
units per micron, which rounds the GDS to 10 nm.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: jeffhsu3 <jeffhsu3@gmail.com>
There was a problem hiding this comment.
Code Review
This pull request introduces a new Tcl utility ring_offsets.tcl with the asap7_ring_core_offsets procedure to dynamically calculate core offsets for PDN rings on routing tracks. It also updates several PDN grid strategy files (BLOCKS_grid_strategy.tcl, BLOCK_grid_strategy.tcl, and grid_strategy-M1-M2-M5-M6.tcl) to use this utility and adjusts various stripe widths, pitches, and connection parameters to ensure proper grid alignment. However, there is a critical issue in both BLOCKS_grid_strategy.tcl and BLOCK_grid_strategy.tcl where the horizontal and vertical layer arguments passed to asap7_ring_core_offsets are swapped, which will result in incorrect offset calculations and off-grid wires.
🔍 QoR checkMetrics reflect the PR merge build — i.e. what will land on the target branch. Commit 62 design(s) checked — 3 with regression(s), 0 without a comparable baseline. ❌ asap7/cva6 base — 1 failing metric(s)
❌ asap7/riscv32i-mock-sram base — 1 failing metric(s)
❌ asap7/swerv_wrapper base — 1 failing metric(s)
|
Summary
The ASAP7 PDN scripts place power vias and stripes where the design rule manual's grid, track and via-width rules cannot hold. This changes all three (
grid_strategy-M1-M2-M5-M6.tcl,BLOCKS_grid_strategy.tcl,BLOCK_grid_strategy.tcl) so the PDN stays on the routing grid and tracks. Part of the divergences listed in #4667.-min_width_layers {M3 M4} -ongrid {M4}.-snap_to_grid, at a fifth of the pitch (1.08 um; 0.432 um at theBLOCKStop level), so the M5 metal is unchanged. M6 straps are-snap_to_gridtoo.ring_offsets.tclcomputes per-side-core_offsetsthat put every ring wire on its layer's tracks. TheBLOCKSring's M5 is 0.408 um (0.504 is over M5's 0.48 um maximum) and its M6 0.288 um, the one widthM6_M5widePWR1p152makes a flush V5 for. TheBLOCKring's M5 is minimum width for the V4s at its corners.Type of Change
Impact
Checked with gdscheck's
asap7openroadsuite: the DRM's rules less the ones the tech LEF does not give the router, so a marker is a DRM violation of a rule the flow could have kept. ORFS's KLayout deck has its grid checks off (OFFGRID = false), so it does not see these.openroadsuite)What remains on
aes-blockis not the PDN: macros placed off the fin/gate grid, and tech LEF via and M8/M9 width gaps (see #4667).The strap pitch is set by IR drop: minimum-width straps at the old 5.4 um pitch raised
aesto 11.6%; 2.16 um gave 5.2%, 1.08 um 2.8%. Denser straps take M5 tracks, so congestion-sensitive designs are worth watching in the QoR check.Verification
gcd,aesandaes-blockrun to completion; detailed routing has 0 DRC errors on all three, andaes-block's VDD/VSS are fully connected (PSM).analyze_power_grid(aes-blockwithPWR_NETS_VOLTAGES/GND_NETS_VOLTAGESset, as its config clears them).gcdandaeswere built withSYNTH_USE_SYN=0:syncurrently writes DEF at 100 units per micron, which rounds the GDS to 10 nm (syn: give the exported block the tech's DEF units OpenROAD#11700). QoR metrics are not affected.Related Issues
#4667
🤖 Co-authored with Claude Code