BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260730T152640Z
LOCATION:Exhibit Hall
DTSTART;TZID=America/Los_Angeles:20260728T173800
DTEND;TZID=America/Los_Angeles:20260728T173800
UID:dac_DAC 2026_sess306_LBR008@linklings.com
SUMMARY:Late Breaking Results: Power Mesh Construction for 3D-IC with Back
 side Power Delivery
DESCRIPTION:Chien-Pang Lu (Intel), Chun-Hao Lai and Iris Hui-Ru Jiang (Nat
 ional Taiwan University), and Chih-Hsiang Yang and Chung-Ching Peng (Intel
 )\n\nA 3D-IC architecture packs a design with enhanced functionality\nand 
 density into a small footprint while improving performance\nand lowering c
 osts. By leveraging through-silicon via (TSV) and\ndie-to-die bump technol
 ogy, power can be efficiently delivered to\nthe top and bottom dies. Moreo
 ver, advanced backside power deliv-\nery technology enables a more streaml
 ined power delivery network\nfor 3D-IC. However, the deployment of TSVs an
 d power/ground\nbumps while achieving fast and accurate verification furth
 er com-\nplicates power delivery network (PDN) construction. To tackle thi
 s\nchallenge, in this work, we collect golden IR drop results from com-\nm
 ercial tools for different configurations and densities on TSVs,\nbackside
  metals, die-to-die bumps and C4 bumps then apply and\ncompare with the ef
 ficient tree-based methods (Random Forest and\nXGBoost) and deep neural ne
 tworks (U-Net and Inception U-Net).\nIn saving runtime, compared to heuris
 tic manual approach, Ran-\ndom Forest can achieve a 5.8𝑋 speedup (includin
 g commercial tool\nrun time). In reducing the power mesh metal ratio, we d
 evise an\nInception U-Net combined with image rotation to generate end-\nt
 o-end IR-drop prediction results for 3D-IC, achieving an MSE of\n0.002/0.0
 05 (bottom/top die), 𝑅2 of 0.97/0.95 (bottom/top die), saving\n7.6% metal 
 usage, and delivering more than a 5% improvement in\nIR-drop results, whic
 h are very close to the golden results reported\nby commercial tools. The 
 experiments are conducted on industrial\nhomogeneous design integrations m
 anufactured by a 3nm process.\nOur results show that our approach can quic
 kly and accurately\npredict IR drop, establish a PDN with small resource u
 sage, and\ngreatly reduce the time for back-and-forth verification.\n\nTra
 ck: Student\n\n
END:VEVENT
END:VCALENDAR