1. 1 Modeling and Optimization of VLSI Interconnect 049031 Lecture 1: Introduction Avinoam Kolodny Konstantin Moiseev

2. 2 Why bother about interconnects?  Chip performance and power depend on interconnect  VLSI Technology has changed: “Device-dominant” → “Interconnect-dominant”  Interconnect has changed: “Parasitic effect” → “primary effect”  What are the reasons for the changes?  Growth in system complexity  Poor scaling of interconnects

3. 3 Process Technology Scaling: Minimum Feature Size Source: Intel, SIA Technology Roadmap SIA: Semiconductor Industry Association 0.01 Feature Size (microns) 0.1 1 10 ’68’71’76’80’84’88’92’96’00’04 ’08 Intel SIA ’14 130nm 90nm 65nm 45nm 32nm 180nm 22nm 14nm 22nm

4. 4 Scaling: Moore’s Law Source: Gordon E. Moore, Cramming more components onto integrated circuits, Electronics, April 19, 1965, p.114 1965 version 1979 version

5. Processor Chips ! ! ! !! 5

6. 6 Organization of Chips  Blocks (cells)  Ports (pins)  Nets (nodes)  Local wires  Global wires

7. 7 Keeping up with Moore’s Law: Rent’s Rule  Abstraction  Hierarchy  Regularity  Design Methodology 100 1000 10 0 10 1 10 2 10 3 10 4 T Terminals per module 1 10 N Transistors in module Rent’s exponent <1

8. 8 Interconnect Length Distribution Source: Shekhar Y. Borkar, Nir Magen - Intel

9. 9 Connectivity and Complexity

10. 10 3-dimensional view of interconnect

11. 11 History 1: Single metal layer Intel 8088 processor When interconnections layers were limited, people preceived Systolic arrays Source: http://www.microscopy.fsu.edu/chipshots/intel

12. 12 History 2: Two layers of metal

13. 13 Interconnect Stack

14. 14 Growing demand for metal layers Source: ITRS 2007

15. 15 45nm (intel)  Cu Wiring  Low-k ILD  Narrow plugs  Stacked vias  Note aspect ratio and wire spacing! M1-M3 M4 M5 M6 M7 M8

16. 16 Metal Processing Single Damascene Dual Damascene ILD Deposition Oxide Trench Etch Metal Fill Metal CMP Oxide Trench / Via Etch Metal Fill Metal CMP

17. 17 The Interconnect Scaling Problem: Transistors get better, Wires get worse

18. 18 Transistor Scaling (ideal)

19. 19 Wire Scaling (ideal)

20. 20 Interconnect time constant

21. 21 Nonuniform Scaling of Wires The idea: Shrink lateral dimensions – save area Keep vertical dimensions – avoid high resistance

22. 22 Non-uniform scaling of wires

23. 23 Interconnect scaling: Wire Capacitance ε 0 = 8.85 x 10 -14 F/cm ε r = 3.9 (SiO 2 )

24. 24 “interconnect scaling – the real limiter” Bohr, IEDM 95 P T T T

25. 25 “interconnect scaling – the real limiter” Bohr, IEDM 95; ITRS 1997 -Unloaded single transistors -Fixed-length wire

26. 26 Criticism on Bohr’s model  Assumed fixed wire length of 1mm  Assumed a fixed (small) transistor size  Ignored the effect of transistor size on circuit speed

27. 27 Review: Why transistors become better with scaling?

28. 28 Review: Why wires become worse with scaling? Local wires: Global wires:

29. 29 Local wires and Global wires  Local wire:  Shrinks in length just like everything else  While transistors become faster, local wire RC remains unchanged (by simple scaling theory)  Global wire:  Goes across the whole chip – does not scale!  Reflects new complexity added to the system!  Global wire RC grows as (wirelength) 2

30. 30 “The bottom of deep submicron” (Sylvester & Keutzer, ICCAD 98)  How to separate gate delay from wire delay?  How strong should the driving gate be? Conclusion: 50K-100K gate blocks are OK for traditional design flow.

31. 31 “The future of wires” (Ho, Mai, Horowitz, Proc. IEEE 2001 )  Local wires scale in performance  Global wires do not  Implications:  CAD tools to improve handling of long wires and “exceptions” (otherwise design productivity will be destroyed)  VLSI architecture must explicitly account for global latencies

32. 32 Delay of global wire is longer than a clock cycle  Time for signal propagation across the die:  Today: 2-3 cycles  In 10 years: ~10 cycles  Fraction of chip reachable in a single cycle:  Today: ~25%  In 10 years: ~1% Fraction of chip reachable in 1 clock cycle Source: Keckler et al. ISSCC 2003

33. 33 Uniprocessor architecture inefficiency “Fred Pollack’s rule”: New microarchitectures take a lot more area for just a little more performance New microarchitecture ~2-3X die area of the last uArchNew microarchitecture ~2-3X die area of the last uArch Provides 1.5-1.7X performance of the last uArchProvides 1.5-1.7X performance of the last uArch Die Area Performance Global interconnect delay is part of this problem!

34. 34 Future of VLSI architecture (Dally 1999, Horowitz 2001)  System requirement: Power-efficient performance growth  Implications:  Chip Multi Processors (CMP)  Thread-Level Parallelism (TLP)  Local memories  Explicit communication M P M P PP PP M M P M P

35. 35 The infamous growth in processor power Source: Intel (Sery, Borkar and De, DAC 2002)

36. 36 Interconnect power  Dynamic power P = ΣAF i C i V 2 f  Definition: Interconnect-Power Dynamic power consumption due to interconnect capacitance switching

37. 37 Interconnect Power in the Banias chip Total dynamic powerInterconnect power Source: Magen et al., SLIP 2004

38. 38 Additional issues with wires  Delay  Power  Noise  Reliability  Cost

39. 39 Noise: on-chip communication will become unreliable  Crosstalk  Capacitive and inductive  Power supply noise  Reduced threshold voltage and margins  Timing errors  Synchronization failures between different clock domains  Soft errors : logic upsets  100% Reliable transmission over long wires will not be possible  Need error correction mechanisms

40. 40 Capacitive Crosstalk

41. 41 Capacitive Crosstalk Noise

42. 42 The future of design flow?  J. Cong : “Interconnect-centric design” Proc. IEEE 2001

43. 43 Summary  Growing system complexity requires more wires  Poor scalability of global wires  Delay, Power, Noise,…..  Interconnect-centric design is desirable