Chip Multicore Processors - Tutorial 2: 2.1: Frequency and Voltage Scaling, Amdahl's Law
Chip Multicore Processors - Tutorial 2: 2.1: Frequency and Voltage Scaling, Amdahl's Law
(1)
A colleague suggests to combine and C to a processor implementation-dependent coecient . For the given processor 0 = 8, 6 104 W/(MHzV2 ) applies. What is the power dissipation and power density? How many instructions can be executed per second? b) You remember a simple estimation from the lecture, that describes how you can scale the supply voltage and frequency to save power. You therefore plan with four processor cores and simplify by assuming the software can be perfectly parallelized. How does the power change? c) Your colleague is proud of your math skills, but reminds you the processor cannot run with the supply voltage. Instead a miminal voltage of 0, 9 V is required. At this voltage you will get something around maximum 1, 4 GHz. Redo you calculations. d) After that you need a coee. In the kitchen you meet a guy from another working group. They work on a power-optimized version of the processor. After a few minutes you nd a mail containing their specication: Vdd = 0, 9 V, f = 800 MHz with a reduced area of 4.5 mm2 . You ask for you new invented value of this implementation and your colleague coarsely estimates it as ab zu 1 = 7, 71 104 W/(MHzV2 ). Calculate the power for this processor variant. e) These calculations start to become fun to you. What is still missing is a real world example. After some phone calls to systems and software-guys you receive data regarding a multimedia application of a custones, which executes 8, 4 109 instructions per second on the processor. Lucky you, a colleague already analyzed the source code and did some proling. He found 14,3% of the application are sequential. You shortly discuss the application and come to the conclusion the remaining parallel share can be executed on the four processor cores. Which of
2 above congurations suits the application? Calculate the average utilization of the processor cores. f) You rethink the whole setup. It might be worth running the three cores that do not run the sequential part with a slower clock frequency. The sequential parts needs to be run faster then. Find a system of the available cores that matches these conditions. After a long rst working day you feel satised and go home.