Routing and placement in semicondoctor vlsi .pdf
- 1. January 25, 2019 Disclaimer: This course was prepared, in its entirety, by Adam Teman. Many materials were copied from sources freely available on the internet. When possible, these sources have been cited; however, some references may have been cited incorrectly or overlooked. If you feel that a picture, graph, or code example has been copied from you and either needs to be cited or removed, please feel free to email adam.teman@biu.ac.il and I will address this as soon as possible. Digital VLSI Design Lecture 9: Routing Semester A, 2018-19 Lecturer: Dr. Adam Teman
- 2. © Adam Teman, 2018 Routing: The Problem • Scale • Millions of wires • MUST connect them all • Geometric Complexity • Basic starting point – grid representation. • But at nanoscale – Geometry rules are complex! • Also, many routing layers with different “costs”. • Electrical Complexity • It’s not enough to just connect all the wires. • You also have to: • Ensure that the delays through the wires are small. • Ensure that wire-to-wire interactions (crosstalk) doesn’t mess up behavior. 2
- 3. © Adam Teman, 2018 Problem Definition • Problem: • Given a placement, and a fixed number of metal layers, find a valid pattern of horizontal and vertical wires that connect the terminals of the nets. • Input: • Cell locations, netlist • Output: • Geometric layout of each net connecting various standard cells • Two-step process • Global routing • Detailed routing 3 • Objective • 100% connectivity of a system • Minimum area, wirelength • Constraints • Number of routing layers • Design rules • Timing (delay) • Crosstalk • Process variations
- 4. © Adam Teman, 2018 Lecture Contents
- 5. Routing Algorithms 5 1 Intro 2 Routing Algorithms This section is heavily based on Rob Rutenbar’s “From Logic to Layout”, Lecture 11 from 2013. For a better ☺ and more detailed explanation, do yourself a favor and go see the original! 2 Routing in Practice
- 6. © Adam Teman, 2018 Grid Assumption • Despite the complexity of nanoscaled routing, we will use a grid assumption and add the complexity in later. • Layout is a grid of regular squares • A legal wire is a set of connected grid cells through unobstructed cells. • Obstacles (or blockages) are marked in the grid. • Wires are strictly horizontal and vertical (Manhattan Routing) • Paths go • North/South • East/West. Obstacle S T Source Target
- 7. © Adam Teman, 2018 Maze Routers • Also known as “Lee Routers” • C. Y. Lee, “An algorithm for path connections and its applications” 1961 • Strategy: • Route one net at a time. • Find the best wiring path for the current net. • Problems: • Early wired nets may block path of later nets. • Optimal choice for one net may block others. • Basic Idea: • Expand → Backtrace → Cleanup 7
- 8. © Adam Teman, 2018 Maze Routing: Expansion • Start at the source. • Find all paths 1 cell away. • Continue until reaching the target. • We approach the target with a “wavefront” • We found that the shortest path to the target is 6 unit steps. 8 S T 1 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4 5 5 5 5 5 5 5 6 6 6 6 6 6
- 9. © Adam Teman, 2018 Maze Routing: Backtrace & Cleanup • Backtrace: • Follow the path lengths backwards in descending order. • This will mark a shortest-path to the target. • However, there may be many shortest paths, so optimization can be used to select the best one. • Cleanup • We have now routed the first net. • To ensure that future nets do not try to use the same resources, mark the net path from S to T as an obstacle. 9 1 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4 5 5 5 5 5 5 5 6 6 6 6 6 6 S T
- 10. © Adam Teman, 2018 Maze Routing: Blockages • How do we deal with blockages? • Easy. Just “go around” them! To summarize: • Expand: • Breadth-first-search (BFS) to find all paths from S to T in path-length order. • Backtrace: • Walk shortest path back to source. • Cleanup: • Mark net as obstacle and erase distance markers. 10 S T 1 2 2 2 3 3 3 3 4 4 4 4 5 5 5 6 6
- 11. © Adam Teman, 2018 Multi-Point Nets • How do we go about routing a net with multiple targets? • Actually, pretty straightforward. • Start with our regular maze routing algorithm to find the path to the nearest target. • Then re-label all cells on found path as sources, and re-run maze router using all sources simultaneously. 11 T S T 1 2 2 2 2 3 3 3 3 3 3 3 S 4 4 4 4 4 4 4 4 S S S
- 12. © Adam Teman, 2018 Multi-Point Nets • How do we go about routing a net with multiple targets? • Actually, pretty straightforward. • Start with our regular maze routing algorithm to find the path to the nearest target. • Then re-label all cells on found path as sources, and re-run maze router using all sources simultaneously. • Repeat until reaching all target points. Note that this does not guarantee the shortest path (=“Steiner Tree”) 12 S T 5 5 5 4 4 4 4 4 4 4 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 S S S 1 1 1 1
- 13. © Adam Teman, 2018 Multi-Layer Routing • Okay, so what about dealing with several routing layers? • Same basic idea of grid – one grid for each layer. • Each grid box can contain one via. • New expansion direction – up/down. 13
- 14. © Adam Teman, 2018 Multi-Layer Routing 14 T 8 8 8 8 8 8 8 8 7 7 7 7 7 7 7 6 6 6 6 7 7 7 6 6 6 6 6 5 5 5 5 5 5 2 2 2 2 3 3 3 3 3 3 S 1 Metal 1 Metal 2 V V V V 4 4 4 4 4 8 8 5
- 15. © Adam Teman, 2018 Non-Uniform Grid Costs • But we know that vias have (relatively) high resistance. • Shouldn’t we prefer to stay on the same metal layer? • We also prefer Manhattan Routing • Each layer is only routed in one direction. • A “turn” requires going through a via or a “jog” should be penalized. • Is there a way to prefer routing in a certain layer/direction? • Yes. • Let’s introduce non-uniform grid costs. 15
- 16. © Adam Teman, 2018 Multi-Layer Routing 16 T 27 27 17 26 16 16 15 15 6 13 4 S 11 2 Metal 1 Metal 2 V V V V 5 14 25 25 1 Horizontal Vertical Cost of Via = 10 12 3 13 16 25 14 23 24 15 35 26 35 26 26 Cost of wrong way route = 10 Cost of wrong way route = 10
- 17. © Adam Teman, 2018 How do we implement this? • Grids are huge. • Assume 1cm X 1cm chip. • Assume 100 nm track • Assume 10 routing layers • That is 1010 (100 billion) grid cells! • We need a low cost representation • Only store the wavefront. • Remember which cells have been reached, at what cost, and from which direction. • Use Dijkstra’s algorithm to find the cheapest cell first. • Store data in a heap. 17 All of this is hard! Use many different heuristics: • Which net to route first • Bias towards the right direction • How to go about fixing problems • etc., etc., etc.
- 18. © Adam Teman, 2018 Divide and Conquer: Global Routing • To deal with a big chip, we make our problem smaller • Divide the chip into big, course regions • e.g., 200 X 200 tracks each. • These are called GBOXes. • Now Maze Route through the GBOXes 18
- 19. © Adam Teman, 2018 Divide and Conquer: Global Routing • Global routing takes care of basic congestion. • Balances supply vs. demand of routing resources. • Generates regions of confinement for the wires. • Detailed routing decides on the exact path. 19 GBOXes have a dynamic cost according to how congested they are!
- 20. Routing in practice 20 1 Intro 2 Routing Algorithms 2 Routing in Practice
- 21. © Adam Teman, 2018 Layer Stacks • Metal stacks are changing (and growing) 130 nm 90 nm 65 nm 45 nm 32 nm M1 M2 M3 M4 M5 B1 B2 M1 M2 M3 M4 B1 B2 B3 E1 E2 M1 M2 M3 M4 B1 B2 C1 C2 B3 E1 U1 U2 M1 M2 M3 M4 B1 B2 B3 E1 E2 M5 W1 W2 M1 M2 M4 M5 M6 M3 Representative layer stacks for 130 nm - 32 nm technology nodes Intel 45nm 8 metal stack UMC 6 metal stack
- 22. © Adam Teman, 2018 Global Route • Divide floorplan into GCells • Approximately 10 tracks per layer each. • Perform fast grid routing: • Minimize wire-length • Balance Congestion • Timing-driven • Noise/SI-driven • Keep buses together • Also used for trial route • During earlier stages of the flow 22 Horizontal routing capacity = 9 tracks Vertical routing capacity = 9 tracks X X Y Y
- 23. © Adam Teman, 2018 Congestion Map • Use congestion map and report to examine design routability 23 Congestion map Congestion Report # Routing #Avail #Track #Total %Gcell # Layer Direction Track Blocked Gcell Blocked # -------------------------------------------------------------- # Metal 1 H 7607 9692 1336335 62.57% # Metal 2 H 7507 9792 1336335 55.84% # Metal 3 V 7636 9663 1336335 59.51% # Metal 4 H 8609 8691 1336335 52.02% # Metal 5 V 5747 11551 1336335 56.39% # Metal 6 H 5400 11899 1336335 55.09% # Metal 7 V 1831 2486 1336335 55.30% # Metal 8 H 2415 1903 1336335 43.85% # -------------------------------------------------------------- # Total 46753 56.99% 10690680 55.07% # # 589 nets (0.47%) with 1 preferred extra spacing.
- 24. © Adam Teman, 2018 Detailed Route • Using global route plan, within each global route cell • Assign nets to tracks • Lay down wires • Connect pins to nets • Solve DRC violations • Reduce cross couple cap • Apply special routing rules • Flow: • Track Assignment (TA) • DRC fixing inside a Global Routing Cell (GRC) • Iterate to achieve a solution (default ~20 iterations) 24 Detailed Route Boxes Solve shorts Notch Spacing Notch Spacing Thin&Fat Spacing Min Spacing
- 25. © Adam Teman, 2018 Timing-Driven Routing • Optimize critical paths • Route some nets first • Most routing freedom at start • Use shortest paths possible • Net weights • Order of routing (priorities: e.g., Default : Clocks 50, others 2) • Wire widening • Reduce resistance • If you have a congested design you may need to set the timing driven effort to “low” • Beware when changing default options
- 26. © Adam Teman, 2018 Signal Integrity (SI) • Signal Integrity during routing is synonymous with Crosstalk. • A switching signal may affect a neighboring net. • The switching net is called the Aggressor. • The affected net is called the Victim. • Two major effects: • Signal slow down • When the aggressor and victim switch in opposite directions. • Signal speed up • When the aggressor and victim switch in the same direction. 26 Aggressor Victim Delay net 1 net 2 Speed Up
- 27. © Adam Teman, 2018 • Infinite Window Analysis • An infinite noise window applies the maximum delay due to crosstalk during timing analysis. • This model was sufficient for older (pre-90nm) technologies, but became too severe with the growing sidewall capacitances at scaled nodes. • Propagated Noise Analysis • Min/Max vectors are propagated through the design to create a transition window for all aggressors in relation to a certain victim. • Noise is only applied at the overlap of the two windows to determine the worst case noise bump. SI Multi-Aggressor Timing Analysis 27 Net X aggressor Net Y aggressor Crosstalk noise from X Crosstalk noise from Y Propagated noise from B Worst-case combination of noise bumps on net C Noise bumps on net C
- 28. © Adam Teman, 2018 Signal Integrity - Solutions • Crosstalk Prevention • Limit length of parallel nets • Wire spreading • Shield special nets • Upsize driver or buffer 28 Extra space Spacing Shielding Same layer (H) Adjacent layers (V) Grounded shields Net Ordering Critical Nets Wire Spreading Adding shielding Inserting buffer Upsizing victim driver Increasing wire spacing
- 29. © Adam Teman, 2018 Design For Manufacturing • During route, apply additional design for manufacturing (DFM) and/or design for yield (DFY) rules: • Via reduction • Redundant via insertion • Wire straightening • Wire spreading 29 Wire straightening (reduce jogs) Avoid asymmetrical contacts
- 30. © Adam Teman, 2018 DFM: Via Optimization • Post-Route Via Optimization, includes: • Incremental routing for the minimization of vias. • Replacement of single vias with multi-cut vias. • These operations are required for: • Reliability: • The ability to create reliable vias decreases with each process node. If a single via fails, it creates an open and the circuit is useless. • Electromigration: • Electromigration hazards are even more significant in vias, which are essentially long, narrow conductors. 30 extra vias added
- 31. © Adam Teman, 2018 DFM: Wire Spreading • Wire spreading achieves: • Lower capacitance and better signal integrity. • Lower susceptibility to shorts or opens due to random particle defects. 31 Center of conductive defects within critical area – causing shorts Center of non-conductive defects within critical area – causing opens Center of conductive defects outside critical area – no shorts Center of non-conductive defects outside critical area – no opens Critical Areas Metal 3 + + + + + + + + • High-density critical area • High probability of yield-killing defect • Density reduced, yield risk reduced • No timing impact!
- 32. © Adam Teman, 2018 Routing in Innovus/Encounter • The detailed routing engine used by Innovus/Encounter is called “NanoRoute” • NanoRoute provides concurrent timing-driven and SI-driven routing. • In addition, it can perform multi-cut via insertion, wire widening and spacing. • The commands for running a route with NanoRoute are: • Following detailed route wire optimization and timing optimization: 32 set_db route_design_with_timing_driven true set_db route_design_with_si_driven true route_design set_db route_design_with_timing_driven false set_db route_design_detail_post_route_spread_wire true set_db route_design_detail_use_multi_cut_via_effort high route_design –wire_opt set_db route_design_with_timing_driven true opt_design –post_route –setup -hold
- 33. © Adam Teman, 2018 Routing in Innovus/Encounter • To achieve a high percentage of multi-cut vias: • To check your design after routing: • To perform incremental routing (ECO routing): set_db route_design_concurrent_minimize_via_count_effort high set_db route_design_detail_use_multi_cut_via_effort high report_route; # provide routing statistics report_wires; # provides wire statistics including wirelength time_design –post_route; # check timing after routing check_drc; # Run a DRC check – in new techs: “verify_drc” check_connectivity; # Run an LVS check set_db route_design_with_eco true route_global_detail
- 34. The Chip Hall of Fame • With RISC processors a central part of our computing lives today, we should really thank the revolution of the • Scalable Processor Architecture • Release date: 1987 SPARC v7 32-bit Architecture • The first major commercial RISC processor taking Patterson’s ideas at Berkeley into a product. • “SPARC will take Sun from a $500M/year company to a $1B/year company”. • The first SPARC powered the Sun-4 workstations, which made Sun a $1B/year company. • Terminated in 2017 by Oracle, but now part of Fujitsu. 2017 Inductee to the IEEE Chip Hall of Fame Photo: Mark Richards Photo: Wikipedia SPARCstation 1+ “pizzabox”
- 35. © Adam Teman, 2018 Main References • Rob Rutenbar “From Logic to Layout” 2013 • Synopsys University Courseware • IDESA • Kahng, et al. “VLSI Physical Design: From Graph Partitioning to Timing Closure” – Chapter 6 35