Generated Liberty sets input-pin max_transition (0.04 ns) below the PDK's achievable slew

[cpehle]

Summary

For the OpenRAM-generated sky130 SRAM sky130_sram_2kbyte_1rw1r_32x512_8, the
output Liberty declares max_transition : 0.04 (40 ps) on the input pins
(address and write-mask buses), while the same library's default_max_transition
is 0.5 (500 ps).

40 ps is tighter than sky130 standard cells can physically drive, so any
timing-driven P&R/STA flow that enforces slew limits treats the design as having
an infeasible constraint and stops.

Details

• Macro / config: sky130_sram_2kbyte_1rw1r_32x512_8
  (macros/sram_configs/sky130_sram_2kbyte_1rw1r_32x512_8.py), corner
  TT_1p8V_25C.

• The timing LUTs use the input-transition axis index_1("0.00125, 0.005, 0.04")
  — i.e. inputs are characterized only up to 0.04 ns.

• The input buses then carry that maximum as their max_transition:

  library (sky130_sram_2kbyte_1rw1r_32x512_8_TT_1p8V_25C_lib) {
      time_unit : "1ns" ;
      default_max_transition : 0.5 ;
      ...
      bus (addr0)  { direction : input; max_transition : 0.04; ... }
      bus (wmask0) { direction : input; max_transition : 0.04; ... }
      ...
  }
  

  (three input buses in total) — 12.5x tighter than the file's own
  default_max_transition.

Downstream effect

A timing-driven flow rejects it. With OpenROAD (global_placement -timing_driven
→ resizer slew pre-check):

[ERROR RSZ-0090] Max transition time from SDC is 0.040ns.
Best achievable transition time is 0.098ns with a load of 0.00pF

OpenROAD's check is correct — the smallest sky130 buffer cannot produce a
transition below ~0.098 ns even at zero load — so the 0.04 ns input constraint is
physically unmeetable. The value originates in the Liberty, not the tool
(reproduced with OpenROAD @ 51c9638, but independent of tool version).

Question / suggestion

• Is the input-transition characterization range (max 0.04 ns here) the intended
  default? For a macro whose inputs are driven by ordinary standard-cell logic,
  the characterized input slew should span realistic drive (hundreds of ps), or
  at minimum should not be tighter than the target PDK's achievable transition.
• Would it make sense to (a) widen the default input-slew list used for
  characterization, and/or (b) avoid emitting a per-pin max_transition that is
  tighter than the library's own default_max_transition?

Happy to attach the full .lib or the config if useful.

Environment

• sky130 PDK, ngspice Liberty characterization.
• The same views are also redistributed via efabless/sky130_sram_macros.