Toward Quantifying the IC Design Value of Interconnect Technology Improvement Tuck-Boon Chan, Andrew B. Kahng, Jiajia Li VLSI CAD LABORATORY, UC San Diego http://vlsicad.ucsd.edu UC San Diego / VLSI CAD Laboratory Outline Motivation Related Work Our Framework Experiments and Results Conclusion -2- Outline Motivation Related Work Our Framework Experiments and Results Conclusion -3- Motivation Wire delay increases with technology scaling Improvement of BEOL both important and expensive Issue 1: no systematic quantification of ROI from BEOL improvement Issue 2: unclear whether BEOL improvement benefits can be leveraged by EDA tools Goals: – A framework to quantify BEOL improvement values guide BEOL technology investment and targets – Assess EDA tools’ ability to leverage improved BEOL = potential “EDA gap” -4- Focus of Our Work Product quality comes from interaction among design, BEOL technology, EDA tool We focus on interaction between BEOL and EDA This work EDA tool BEOL Technology Design -5- Outline Motivation Related Work Our Framework Experiments and Results Conclusion -6- Related Work Studies of DRAM or simple logic circuits, not at chip-level Ignores interaction between BEOL technology, EDA tool [Li01] – DRAM performance improvements from low-k [Kapur02] – R, C impacts on signal, power [Bamal06] – Performance, energy comparison studies with different interconnect technologies Focus on variation, not future BEOL improvements [Jeong10] – Chip-level impacts of interconnect variation due to double-patterning -7- Outline Motivation Related Work Our Framework Experiments and Results Conclusion -8- Our Framework Original BEOL files Hypothetical RC reductions Modified BEOL files Designs Circuit implementation flow (synthesis, place and route) Circuit implemented with original BEOL Circuit implemented with modified BEOL Timing and power analysis 1. Modify BEOL files to model R, C reductions in future technologies – Modify ITF files – Use Synopsys StarRC to convert ITF to TLUplus files 2. Design implementation (RTL-to-layout and signoff) with original and modified BEOL files 3. Run timing, power analysis -9- Testbed Designs: {aes_cipher, des_perf, mpeg2, pci_bridge32} from OpenCores x {fast, slow} clock periods Technology: TSMC 45nm, LVT and HVT 20SOC and below can be very different SP&R: Synopsys Design Compiler + IC Compiler – Execute each P&R run three times denoising Timing and power analysis: Synopsys IC Compiler Signoff: no hold or EM violation, TNS < 30ps Apples-to-apples comparison for design metrics -10- Outline Motivation Related Work Our Framework Experiments and Results Conclusion -11- Expt 1: Impact of R, C Reduction on Power 45% R, C reduction only leads to 8% power reduction R, C reduction improves timing fewer / smaller cells leakage power ↓ (but, only on critical paths) C reduction load cap ↓ net switching power ↓ (but, gate cap dominates) Normalized Power 1.1 min average max linear Power of implementation with original BEOL 1.0 0.9 0.8 0.7 R, C=α%: implementation with α% dielectric constant and metal resistivity w.r.t original BEOL 0.6 R, C reduction occurs on M2-M5 R, C=55% R, C=70% R, C=85% R, C=100% -12- Impact of R, C Reduction on Area R, C reduction leads to little improvement in area Tool uses Vt swapping to exploit improved timing – Same footprint of LVT and HVT cells same post-opt area Optimization methodology of EDA tools affects value extracted from improved BEOL 1.1 Normalized Area min average max linear Area of implementation with original BEOL 1.0 0.9 0.8 0.7 0.6 R, C=55% R, C=70% R, C=85% R, C=100% -13- Expt 2: Reduction in R vs. in C In this experiment, C reduction offers more benefits – Wire delay ↓ trade timing for power – R reduction improves wire delay – C reduction improves wire delay + load cap R reduction can be critical with high Vdd, temperature Technology R&D might focus more on C reduction Power w/ only R reduction min average max linear 1.0 0.9 0.8 0.7 1.1 Normalized Power Normalized Power 1.1 Power w/ only C reduction min average max linear 1.0 0.9 0.8 0.7 R=55% R=70% R=85% R=100% C=55% C=70% C=85% C=100% -14- Expt 3: R, C Reduction in Advanced Technology Wire delay becomes critical in advanced technologies – Impact of R reduction increases – We model advanced technology = increase R by 8x Benefits of R, C reduction increase in advanced technologies Leakage power Advanced Current 5% 1.0 2% 0.9 0.8 0.7 R=55% C=55% R, C=55% Normalized Total Power Normalized Leakage 1.1 Total power 1.1 Advanced Current 1.0 0.9 0.8 0.7 R=55% C=55% R, C=55% -15- Expt 4: Impact of Layer Selection BEOL improvement incurs high manufacturing cost – What is optimum subset of layers to improve under cost limits? – Flexible BEOL = subset of layers is selectively improved – Inappropriate selection of R, C-reduced layers is suboptimal Guideline: reduce R, C on adjacent and highly utilized layers Small difference between different layer selections – Tools’ ability to leverage the improved BEOL layers? 1.1 Normalized Power min average RC-reduced layers are far from each other max 1.0 RC reduction has more benefit on highly utilized layers 0.9 0.8 {2,3,4,5} {2,3} {2,4} {2,5} {3,4} {3,5} Layers with improved BEOL {4,5} -16- Tools’ Exploitation of R, C Reduction Assessment flow 1. Implement designs with both original and improved BEOL 2. Run timing and power analysis with improved BEOL 3. Compare frequency, power Preliminary results show tool can leverage R, C reduction – Case 1 might be misguided during optimization 12 13 Reduced R, C on M2, M5 12 11 Power (mW) Power (mW) Reduced R, C on M3, M4 10 9 11 10 9 8 8 800 850 900 950 Frequency (MHz) 1000 1050 800 850 900 950 Frequency (MHz) 1000 Case 1: Implementation with original BEOL, analyzed with modified BEOL Case 2: Implementation with modified BEOL, analyzed with modified BEOL 1050 -17- RC-Awareness in EDA Tools A “smart” router should be aware of improved BEOL – Route setup critical paths on layers with small R, C – Route hold critical paths on layers with large R, C ∆wire distribution (of layer x) = %wire on layer x - %wire on layer x w/ improved BEOL w/ original BEOL Assessment: – Implement designs with flexible BEOL – Check ∆wire distribution of layers for setup- and hold-critical nets -18- Experimental Results Compare ∆wire distribution from current router (bars) and a hypothetical RC-aware router (ovals) – White (Orange) = positive (negative) ∆wire distribution – Same color of bar and dotted oval RC-awareness Router is not fully responsive to BEOL R, C reduction ∆Wire distribution Setup-critical nets 8% 0% -8% 8% 0% -8% 8% 0% -8% 8% 0% -8% Hold-critical nets √ √ √ √ √ √ √ X X √ X √ √ {2,3,4,5} {2,3} {2,4} {2,5} √ √ X {3,4} {3,5} Layers with reduced RC {4,5} 8% 0% -8% 8% 0% -8% 8% 0% -8% 8% 0% -8% √ X X X √ X X X M2 X X √ X X {2,3,4,5} {2,3} {2,4} {2,5} {3,4} M3 X M4 √ X M5 {3,5} Layers with reduced RC {4,5} -19- Outline Motivation Related Work Our Framework Experiments and Results Conclusion -20- Conclusion Framework to quantify impact of interconnect resistance and/or capacitance reductions on chip-level design metrics Reduction in capacitance gives more benefits than in resistance – R reduction can be critical in wire-delay dominant designs (due to high Vdd, temperature or advanced technology) Capability of EDA tools to leverage improved BEOL has room for improvement Ongoing works – Iso-constraints vs. iso-GDS -21- ISO-GDS Expt Basic tradeoffs to exploit improved BEOL – R, C reduction improved timing Vdd ↓ Power ↓ Vdd reduction 100% 98% Performance requirement + device type Activity factor + nominal voltage + device type 96% 94% 92% 0% 10% Frequency improvement 10% 20% 30% 40% Power reduction R, C reduction 100% 8% 96% 6% 92% 4% Gate-wire balance 2% 0% 0% 10% 20% 30% R, C reduction 40% 88% 84% 0% 10% 20% 30% R, C reduction 40% -22- Conclusion Framework to quantify impact of interconnect resistance and/or capacitance reductions on chip-level design metrics Reduction in capacitance gives more benefits than in resistance – R reduction can be critical in wire-delay dominant designs (due to high Vdd, temperature or advanced technology) Capability of EDA tools to leverage improved BEOL has room for improvement Ongoing works – Iso-constraints vs. iso-GDS – Study impact of interconnect R, C reduction across wide supply voltages – Extend our analyses to M1 and middle-of-line layers -23- Acknowledgments Work supported from Sandia National Labs, Qualcomm, Samsung, NSF, SRC, the IMPACT (UC Discovery) and IMPACT+ centers -24- Thank You! Backup Slides Values of Improved BEOL Question 1: What is overall impact of R and/or C reduction(s) on design metrics? – 45% R, C reduction 8% power reduction, similar area Question 2: Which reductions offer more benefits, in R or in C? – C reduction offers more benefits – R reduction can be critical with high Vdd, temperature Question 3: How will impacts of R, C reduction change in advanced technology nodes? – Benefits of R, C reduction increase in advanced technology Question 4: What is optimum subset of layers to improve under cost limits? – Should reduce R, C on adjacent and highly utilized layers -27-