1. Logic Optimization Mohammad Sharifkhani

2. Reading Textbook II, Chapters 5 and 6 (parts related to power and speed.) Following Papers: –Nose, Sakurai, 2000 –Broderson et. al. 2002 (True Power Min.) –Zyuban, 2002

3. Optimization Only speed has been covered –In most cases, power plays a significant role –Power and delay optimization –Delay constraint, power opt. or reverse

6. Portability

7. Energy density

8. Power sources

9. Energy model

10. Dynamic Power Dissipation Energy/transition = C L * V dd 2 Power = Energy/transition *f =C L * V dd 2 * f Need to reduce C L, V dd, andf to reduce power. VinVout C L Vdd Not a function of transistor sizes!

11. Modification for Circuits with Reduced Swing

12. Node Transition Activity and Power

13. Alpha-power based delay model

14. Optimization Goals When both power and delay are important, how can we optimize? –How about minimization of Energy x Delay? –What if one is more important? –What are the variables in design optimization? Usually the designs are either power or delay (speed) limited Can we optimize for the highest speed and reduce the VDD until we get to the desired speed?

15. Performance Optimization

16. Problem: Power minimization for a given speed Lower power  Lower VDD Lower VDD  A slower design A slower design  Lower Vth Lower Vth  Higher leakage Higher leakage  Higher power Goal: –Optimum VDD, Vth (for a given delta Vth)

17. Problem: Power minimization for a given speed Optimal Vth, VDD (given process parameters including variations) Step 1: calculation of delay, energy vs. design variables Ld is logic depth.

18. Given td and Ld, Vthmax can be obtained based on VDD Given ΔVth, Vthmin can be calculated Problem: Power minimization for a given speed When the clock frequency is given, we differentiate P(VDD) with respect to VDD and set the resultant expression to zero  too difficult to solve, except with a lot of appx. The formula of power dissipation can be derived, which is denoted as P(VDD).

19. Using Taylor expansion. Power Minimization Results Assuming typical values for the parameters such that Ns=0.048 (S-factor=80mV/decade and TmaX=400K) and alpha=1.3, P LEAK,max is calculated to be about 30% of the total power

20. Power Minimization Results Calculated under various –a (activity factor) –f (clock freq) Numarical solution of the original eq. is also possible Approximations made to achieve the closed form solution is acceptable

21. Reults Again, equations matches well with the numerical results

22. Results When logic depth changes, the optimal values for VDD and Vth changes as well.

23. Cost function based opt. What if we have multiple design tuning variables? –Vdd, size, threshold voltage, architecture, etc… –Each tuning variable relates Energy to Delay in a certain pattern A simplistic approach: –Design for the highest speed, then reduce Vdd to achieve a constraint (e.g., delay) –Wrong! Goal: –What is the proper setting for the design variable under optimum situation?

24. Cost function based opt. Other cost functions can also be used: Where E 0 and D 0 are lower bounds for a give design η is the Hardware Intensity Optimum is when the cost function is minimized Hence:

25. Cost function based opt. A logic can be designed in a fixed supply voltage with different design variable (size) –Different designs Each design corresponds to a different hardware intensity η –We can re-interpret the design variable (size) as a different η

26. Cost func. Based opt. For a given η,the smallest value of Fc can be found at the intersection between the lowest dashed line (related to that η) and the solid line (related to the actual hardware). Hence, η corresponds to a particular value of the tuning variable (e.g. size). Dashed lines: Each set of dashed line corresponds to a fixed Fc=A for that particular η based on: The solid line is the ED curve when a tuning variable (e.g, size) changes under fixed Vdd. Actual hardware delay, energy relation wrt e.g. size(η) Contours based on Fc function and η. Independent of hardware

27. Cost func. Based opt. For a given η,the smallest value of Fc can be found at the intersection between the lowest dashed line (related to that η) and the solid line (related to the actual hardware). When you are tuning the size  as if you are changing the η. Dashed lines: Each set of dashed line corresponds to a fixed Fc=A for that particular η based on: The solid line is the ED curve when a tuning variable (e.g, size) changes under fixed Vdd.. Actual hardware delay, energy relation wrt e.g. size(η) Contours based on Fc function and η. Independent of hardware

28. The following relationship holds at this point: Given the fact that –We have Cost func. Based opt. Hardware intensity (η) is the ratio of the relative increase in the energy to the relative gain in performance locally achievable through tuning a design variable (e.g., size or logic restructuring) at a give supply voltage. % value of power % value of performance

29. Relationship with supply voltage One can use simulation to find Ev and Dv for a given design under various η Cost func. Based opt.

30. Let’s solve minimal E(v, η) for a limited D(v, η), hence: Hence we have: Cost func. Based opt. What it means? In an optimal design, the

31. Hence, for a given supply voltage v, an optimal η can be achieved such that the relative gain in speed/ relative increase in power for variation of v (Dv/Ev) is equal to the relative gain in speed/ relative increase in power for variation of( η.) Simple words: At the optimal point, the relative sensitivities of E and D to all design variables must be the same Cost func. Based opt. The energy-efficient design is achieved when the marginal costs of all the tuning variables are balanced

32. Example: if for a given supply voltage Dv=1 and Ev=2  Other tuning variable (e.g., sizing) must be such that η = 2  1% increase in delay and 2% saving in energy The notion that design for the highest speed and reduce the supply voltage does not provide optimum result Ex: if Dv=1, Ev=2 and sizing is optimal for η = 4 (circuit is not optimal) Cost func. Based opt.

33. Performance Optimization In most applications the case is not clear as which cost function has to be minimized There is usually a bound (constrain) on a certain design parameter: –Delay Goal: –What is the methodology of optimization for delay/power constrained design?

34. Performance Optimization

42. Tuning variables

43. Step by step optimization Step1 Find the sensitivity of Energy to various design variables Find the sensitivity of Delay to various design variables Step2 The true power minimization method always exploits the tuning variable with the largest capability for energy reduction. This ultimately leads to the point where the energy reduction potentials of all tuning variables are equalized.

44. Input width Load width

45. Delay time increases the leakage E over one clock cycle

48. In single variable optimization we go along with the variable with the highest potential for energy reduction

52. -From logical effort we know that the delay is minimized when all stage efforts are the same (h eff =4) -Large last stages -Let’s resize the last stages to save power  delay increase (d inc )

53. Trans. Sizing

54. The energy increase due to the VDD raise is compensated by tuning W

58. Consistent with the observations made in the first approach