1. EE4271 VLSI Design Interconnect Optimizations Buffer Insertion

2. Moore’s law Twice the number of transistors, approximately every two years, so double clock frequency accordingly

3. 3 0.18 Source: Gordon Moore, Chairman Emeritus, Intel Corp. 0 50 100 150 200 250 300 Technology generation (  m ) Delay (psec) Transistor/Gate delay Interconnect delay 0.80.50.25 0.15 0.35 Interconnects Dominate This is why Moore’s law is not true anymore.

4. Objectives What have we learned? –Compute circuit delay on wires and gates –Gate delay optimization What are we going to learn? –Interconnect delay optimization: buffer insertion Why reducing delay How to perform it –This is the most important optimization in circuit design

5. 5 0.18 Source: Gordon Moore, Chairman Emeritus, Intel Corp. 0 50 100 150 200 250 300 Technology generation (  m ) Delay (psec) Transistor/Gate delay Interconnect delay 0.80.50.25 0.15 0.35 Why is this trend?

6. A scaling primer Ideal process scaling: –Device geometries shrink by S  = 0.7x) Device delay shrinks by s –Wire geometries shrink by  Unit resistance R/  :  /(ws.hs) = r/s 2 Unit coupling capacitance Cc/  : (hs)/(Ss) Resistance doubled, capacitance roughly unchanged for unit length How about the change in wire length? SGD h w l S ll hh SS ww

7. Technology scaling Global (long) interconnect lengths don’t shrink –Global interconnect link cells far apart Local (short) interconnect lengths shrink by s –Local interconnects link cells nearby

8. Interconnect delay scaling Delay of a wire of length l :  int = (rl)(cl) = rcl 2 (a quadratic function of length) Local interconnects :  int : (r/s 2 )(c)(ls) 2 = rcl 2 –Local interconnect delay unchanged Global interconnects :  int : (r/s 2 )(c)(l) 2 = (rcl 2 )/s 2 –Global interconnect delay doubled – unsustainable! Interconnect delay increasingly more dominant

9. Buffer Insertion For Delay Reduction

10. Elmore Delay for Wire x C unit wire capacitance c unit wire resistance r

11. Elmore Delay for Buffer v C u Driving resistance Input capacitance

12. Elmore Delay for A Circuit Delay =  all Ri  all Cj downstream from Ri Ri*Cj Elmore delay to n1 R(B)*(C1+C2) Elmore delay to n2 R(B)*(C1+C2)+R(w)*C2 R(B) C1 R(w) C2 n1 B n2

13. R Buffers Reduce Wire Delay x/2 cx/4 rx/2 t_unbuf = R( cx + C ) + rx( cx/2 + C ) t_buf = 2R( cx/2 + C ) + rx( cx/4 + C ) t_buf – t_unbuf = RC – rcx 2 /4 x/2 cx/4 rx/2 C C R x ∆t∆t

14. Buffered global interconnects: Intuition Interconnect delay = r.c.l 2 /2 Interconnect delay =  r.c.l i 2 /2 < r.c.l 2 /2 (where l =  l j ) since  (l j 2 ) < (  l j ) 2 (Of course, we need to consider buffer delay as well) l1l1 lnln l3l3 l2l2 l

15. Optimal Buffer Insertion on A Wire Delay before buffer insertion = rcL 2 /2 Assume N identical buffers with equal inter-buffer length l (L = Nl) For minimum delay, L R d – On resistance of inverter C g – Gate input capacitance r,c – unit resistance and capacitance … … l

16. Optimal interconnect delay Substituting l opt back into the interconnect delay expression: Delay grows linearly with L (instead of quadratically)

17. Total buffer count Ever-increasing fractions of total cell count will be buffers –70% in 32nm –25% is widely observed 0 10 20 30 40 50 60 70 80 90nm65nm45nm32nm % cells used to buffer nets clk-buf buf tot-buf

18. ITRS projections

19. Exercise 1 Given a wire of length 10 with r=2, c=2, what is its delay? Given a buffer with R d =10, C g =20, after optimally buffering the wire, what is the delay? What if wire length is 100? Any conclusion?

20. Exercise 2 Relationship with gate sizing –If we can size the buffer, what is the best buffer size? –Let R 0 denote the unit size buffer driving resistance, and C 0 denote the unit size buffer input capacitance. Thus, R d =R 0 /h and C g =C 0 h –What is best h leading to smallest delay?

21. Analogy

22. Advancing technology = period of city expansion, more transistors = larger city Interconnects = streets Buffers = gas stations Signal delay (timing) = time to cross the city Buffer insertion = gas station construction

23. Previous Result is Only Theoretical: Discrete Buffer Locations Candidate buffer locations

24. RAT: Required Arrival Time RAT = 100 Wire delay = 80 AT = 0 RAT = 100 Wire delay = 80 AT = 0 RAT = 20 AT = 80

25. Slack: RAT - AT RAT = 100 Wire delay = 80 AT = 0 RAT = 20 AT = 80 Slack = 20 Minimizing circuit delay = maximizing RAT at driver = maximizing slack at driver

26. Motivation for Problem Formulation RAT = 300 AT = 350 Slack = RAT-AT= -50 RAT = 700 AT = 600 Slack = 100 RAT = 300 AT = 250 Slack = 50 RAT = 700 AT = 400 Slack = 300 slack = -50 slack = 50 Decouple capacitive load from critical path RAT = Required Arrival Time Slack = RAT - AT We need to maximum slack or RAT at driver

27. Timing Driven Buffering Problem Formulation Given –A Steiner tree –RAT at each sink –A buffer type –RC parameters –Candidate buffer locations Find buffer insertion solution such that the slack (or RAT) at the driver is maximized

28. An Example for Buffer Insertion (v 1, 1, 20) 22 v1v1 v1v1 (v 2, 3, 16) r = 1, c = 1 R b = 1, C b = 1 R d = 1 (v 2, 1, 13) v1v1 (v 3, 5, 8) v1v1 (v 3, 3, 9) slack = 6slack = 3 Add wire Insert buffer Add wire Add driver CQ

29. Candidate Buffering Solution Definition Each candidate solution is associated with –v i : a node –c i : downstream capacitance –q i : RAT v i is a sink c i is sink capacitance v is an internal node

30. Van Ginneken’s Algorithm Candidate solutions are propagated toward the source

31. Solution Propagation: Add Wire c 2 = c 1 + cx q 2 = q 1 – rcx 2 /2 – rxc 1 r: wire resistance per unit length c: wire capacitance per unit length (v 1, c 1, q 1 ) (v 2, c 2, q 2 ) x

32. 32 Solution Propagation: Insert Buffer c 1b = C b q 1b = q 1 – R b c 1 C b : buffer capacitance R b : buffer resistance (v 1, c 1, q 1 ) (v 1, c 1b, q 1b )

33. Solution Propagation: Add Driver q 0d = q 0 – R d c 0 R d : driver resistance Pick solution with max slack (v 0, c 0, q 0 ) (v 0, c 0d, q 0d )

34. Exercise (20,400) 2 2 Unit Wire Cap = 5 Unit Wire Res = 3 Buffer C=5, R=1 Perform buffer insertion to maximize the slack at driver 2

35. Exponential Runtime 2 solutions 4 solutions 8 solutions 16 solutions n candidate buffer locations lead to 2 n solutions

36. Solution Pruning Two candidate solutions –(v, c 1, q 1 ) –(v, c 2, q 2 ) Solution 1 is inferior if –c 1 > c 2 : larger load –and q 1 < q 2 : tighter timing

37. LOAD An Analogy - I Faster -> Smaller Delay -> Larger RAT (since RAT = RAT output - Delay) Larger Load -> Larger Capacitance

38. LOAD Faster & smaller load (larger RAT, smaller capacitance): Good Slower & larger load (smaller RAT, larger capacitance): Inferior END An Analogy - II

39. END Who will be the winner? Cannot tell at this moment, so keep both of them. An Analogy - III

40. END Who will be the winner? Cannot tell at this moment, so keep both of them. An Analogy - IV

41. Pruning When Insert Buffer They have the same load cap C b, only the one with max q is kept

42. 42 Generating Candidates (1) (2) (3) From Dr. Charles Alpert

43. 43 Pruning Candidates (3) (a) (b) Both (a) and (b) “look” the same to the source. Throw out the one with the worse slack (4)

44. 44 Candidate Example Continued (4) (5)

45. 45 Candidate Example Continued After pruning (5) At driver, compute which candidate maximizes slack. Result is optimal.

46. 46 Example (20,400) (30,250) (5, 220) (20,400) (30,250) (5, 220) (40, 40) (5, 0) (15,160) (5, 145) Unit Wire Cap = 5 Unit Wire Res = 3 Buffer C=5, R=1 2 2 2

47. 47 Example Cont’d (20,400) (30,250) (5, 220) (40, 40) (5, 0) (15,160) (5, 145) (5,0) is inferior to (5,145). (45,40) is inferior to (15,160) (20,400) (30,250) (5, 220) (15,160) (5, 145) (5,15) (5,70) Pick solution with largest slack, follow arrows to get solution

48. Exercise Without pruning, there will be exponential number of candidate solutions (i.e., given n candidate buffer locations, there will be 2 n solutions). With pruning, how many solutions will we have?

49. Exercise Unit Wire Cap = 1 Unit Wire Res = 1 Buffer C=1, R=1 2 2 (10,40) (8,50) (5,10) (15,40) (7,10) (9,30) (12,20) Continue the following buffer insertion process. Assume that all partial candidate buffering solutions are as shown.

50. Summary Interconnect delay increases with technology scaling Linear interconnect delay with buffer insertion Buffer insertion with candidate buffer locations Pruning for accelerating buffer insertion technique