# ===================================================================== # multi-corner step - MULTI-CORNER power study (real data). # Tools: VCS W-2024.09-SP2-5 (gate-level VCD, w_mult workload) + PrimeTime PX # V-2023.12-SP5-4 (averaged power). Fixed netlists (synthesized at ss for # timing closure); power re-characterized at 3 corners by re-linking the # corner .db (switching activity is corner-independent). # Corners: ss=ss0p75v125c (worst), tt=tt1p05v25c (typical), ff=ff1p16v25c (fast). # Flow: ptpx/run_multicorner.sh + run_ptpx_corner.tcl. # # CONTEXT: this experiment replaced the originally-planned post-layout P&R # (dir A), which is BLOCKED on this machine (no IC Compiler II license; classic # icc_shell missing libQtScript.so.4). See PAPER_PUBLICATION_README.md sec 7.7. # ===================================================================== TABLE - absolute power per (config, corner) [REAL] --------------------------------------------------------------------------- config corner total(mW) dynamic(uW) leakage(mW) clock-net(uW) baseline ss 1.4840 49.6439 1.4340 49.63 baseline tt 16.3000 94.8080 16.2000 94.81 baseline ff 196.4000 1448.0098 194.9000 1449.00 opt1_rtlcg ss 1.1970 23.0138 1.1740 22.66 opt1_rtlcg tt 12.9000 44.6820 12.9000 43.95 opt1_rtlcg ff 164.9000 417.3653 164.5000 416.60 opt2_synthcg ss 1.0500 12.5810 1.0370 12.19 opt2_synthcg tt 11.4000 24.9886 11.4000 24.18 opt2_synthcg ff 148.0000 272.9586 147.7000 272.10 opt3_both ss 1.1410 31.2270 1.1100 28.75 opt3_both tt 12.2000 60.5780 12.1000 55.51 opt3_both ff 158.2000 592.3260 157.6000 586.80 REDUCTIONS vs baseline (%), by corner [REAL, derived] --------------------------------------------------------------------------- corner cfg total dynamic leakage ss opt1_rtlcg -19.3% -53.6% -18.1% ss opt2_synthcg -29.2% -74.7% -27.7% ss opt3_both -23.1% -37.1% -22.6% tt opt1_rtlcg -20.9% -52.9% -20.4% tt opt2_synthcg -30.1% -73.6% -29.6% tt opt3_both -25.2% -36.1% -25.3% ff opt1_rtlcg -16.0% -71.2% -15.6% ff opt2_synthcg -24.6% -81.1% -24.2% ff opt3_both -19.5% -59.1% -19.1% FINDINGS (honest): 1. In this 32nm educational PDK, LEAKAGE dominates total power at EVERY corner (96-99%). Leakage rises steeply ss->tt->ff (baseline 1.43 -> 16.2 -> 194.9 mW) because the supply voltage (0.75 -> 1.05 -> 1.16 V) and fast process outweigh the lower temperature. So a "typical corner" does NOT make the design dynamic-dominated -- the opposite is true. 2. NEVERTHELESS clock gating reduces TOTAL power at every corner, robustly: ICG (opt2) cuts total power -29% (ss), -30% (tt), -25% (ff). 3. Because total is leakage-dominated, that total-power benefit tracks the LEAKAGE reduction (opt2: -28%/-30%/-24%), which clock gating delivers via reduced area / simpler register clouds (shared ICG cells replace per-register feedback muxes; baseline area 19474 -> opt2 17745 um^2 at ss). 4. The DYNAMIC-power reduction is large and corner-robust (opt2: -75% ss, -74% tt, -81% ff). It dominates ENERGY in duty-cycled/active operation even though it is a small fraction of average total power in this leakage-heavy regime. 5. Ordering is corner-robust: ICG (opt2) is best at all corners; opt3 (both) does not beat opt2. TAKEAWAY: Clock gating gives a corner-robust ~25-30% total-power and 54-81% dynamic-power reduction. In a leakage-dominated regime the total-power win comes from the area/leakage reduction that gating brings, not from the (large) dynamic saving; the dynamic saving matters for active-mode energy. This argues for pairing clock gating with leakage-reduction techniques (power gating / multi-Vt) when leakage dominates. Reproduce: python3 sim/scripts/gen_multi_workloads.py # makes pmem_w_mult.mem bash ptpx/run_multicorner.sh # -> reports/multicorner_power.csv