Recap of Lecture 1 and Homework Check
- Architecture = ISA + organization + hardware, designed together.
- Five classes: PMD, desktop, server, WSC, embedded - each with its own critical concerns.
- Four kinds of parallelism: ILP, DLP (vector/SIMD/GPU), TLP, RLP.
- Designer's task: maximize performance and energy efficiency within cost, power and availability limits.
Homework check
- 1. Downtime per year at 99.95% availability?
- 2. Your phone: which ISA, how many cores, what clock speed?
- 3. Was MARS installed for Practical 1?
Today's Plan (50 minutes)
0-4
Recap and homework check Lecture 1
4-7
Plan and learning outcomes
7-17
Module 1: Technology trends Moore's law, DRAM, flash, disk
17-24
Module 2: Bandwidth vs latency why one improves faster
24-36
Module 3: Power and energy formulas and worked examples
36-42
Module 4: End of Dennard scaling power wall, multicore, saving energy
42-50
Activity, recap quiz and homework four short numericals
Learning Outcomes (SLOs)
By the end of this lecture you will be able to:
SLO 2.1
Explain Moore's law and the growth trends of logic, DRAM, flash and disk technology.
→ CLO1 · C2 Understand
SLO 2.2
Compare improvements in bandwidth and latency and explain why bandwidth grows faster.
→ CLO1 · C2 Understand
SLO 2.3
Calculate dynamic energy, dynamic power and static power, and the effect of voltage and frequency scaling.
→ CLO2 · C3 Apply
SLO 2.4
Explain why Dennard scaling ended and how designers save energy (DVFS, clock gating, turbo).
→ CLO1 · C2 Understand
- CLO1: Describe the principles of computer design, including technology, cost and dependability trends.
- CLO2: Measure and report computing performance using benchmarks, metrics, the CPU performance equation and Amdahl's law.
Module 1: Technology Trends
A computer takes years to design, so it must be designed for the technology of the future
Moore's Law: Transistor Count Doubles
2,300
16 billion
2.16 yrs
Intel 4004 (1971)
Apple M1 (2020)
per doubling
transistors
transistors
worked out below
- Growth = 16x10^9 / 2,300 = 6.96 million = 2^22.7
- Years per doubling = 49 years / 22.7 doublings = 2.16 years
Careful: Moore's law is about the number of transistors, not clock speed. Speed gains came from how designers used those transistors.
How Fast Each Technology Improves
Cost per bit: flash is 15-20x cheaper than DRAM; disk is 15-25x cheaper than flash. Source: Hennessy and Patterson (2012), §1.4.
Design lesson: A chip planned today ships in 3-5 years, so it is designed for the technology of that future year.
Smaller Transistors: Faster Switches, Slower Wires
Transistors get better
Wires do not keep up
- Feature size (smallest width on the chip) fell from about 10 µm (1971) to a few nm today.
- More transistors fit, and each switches faster with less energy.
- Thinner wires have more resistance, so a signal crossing the chip is not much faster.
- Wire delay is now a key limit on clock speed and chip layout.
Example: A wider GT Road lets more cars pass (more transistors), but the distance from Faisalabad to Lahore stays the same (wire delay).
Module 2: Bandwidth vs Latency
Throughput improves much faster than delay
Bandwidth and Latency: Two Different Questions
Bandwidth (throughput)
Latency (response time)
- How much data or work is done per second.
- Units: MB/s, GB/s, transactions/s.
- Motorway: more lanes = more cars per hour.
- How long one task takes from start to finish.
- Units: ns, ms, seconds.
- Motorway: the travel time Lahore to Islamabad does not fall with more lanes.
Exam tip: A Daewoo bus (50 seats) has higher bandwidth than a car (4 seats), even if the car has lower latency.
Why Bandwidth Improves Faster
- Rule of thumb (Hennessy and Patterson):
- bandwidth gain >= (latency gain)^2
- Example: latency improves 4x
- => bandwidth improves >= 4^2 = 16x
Why?
- Bandwidth can be bought with parallelism: wider buses, more banks, more lanes.
- Latency is limited by physics: distance and signal speed.
- Designers hide latency with caches, prefetching and overlap (pipelining).
- Example: home internet download speed rises a lot over the years; the ping to a nearby server falls much less.
Module 3: Power and Energy
The biggest limit on computer design today
Power vs Energy: The Electricity Bill
Power = rate (watts)
Energy = total (joules, kWh)
Decides heat and cooling. A 1,000 W iron gets hotter than a 100 W fan.
Decides battery life and the bill. 1 kWh = 1 'unit' on the electricity bill.
Dynamic Energy and Dynamic Power
- Energy (one transition) = 1/2 x C x V^2
- Energy (pulse 0->1->0) = C x V^2
- Power (dynamic) = 1/2 x C x V^2 x f
- C = capacitive load, V = voltage,
- f = switching frequency
Key facts
- Voltage is squared: the strongest lever.
- Frequency changes power, not energy per operation.
- C grows with transistors and wire length.
- Example: C = 20 nF, V = 1.0 V, f = 3 GHz
- P = 0.5 x 20x10^-9 x 1.0^2 x 3x10^9 = 30 W
Worked Example 1: Lowering Voltage by 15%
- New V = 0.85 V_old, new f = 0.85 f_old
- Energy_new / Energy_old = (0.85)^2 = 0.72 -> 28% less energy
- Power_new / Power_old = (0.85)^2 x 0.85 = 0.61 -> 39% less power
-28%
-39%
energy per operation
power
from V squared only
V squared and lower f
Worked Example 2: DVFS in Your Phone
- Before: V = 1.0 V, f = 2.0 GHz After: V = 0.8 V, f = 1.2 GHz
- Power ratio = 0.8^2 x (1.2 / 2.0) = 0.64 x 0.6 = 0.384 -> 62% less power
- Time ratio = 2.0 / 1.2 = 1.67 (task runs longer)
- Energy ratio = 0.384 x 1.67 = 0.64 -> 36% less energy
Key point: Lowering frequency alone saves power but not energy per task (it just takes longer). Energy falls only when voltage falls.
DVFS (dynamic voltage and frequency scaling) = the chip changes V and f with the workload. Throttling = a hot phone lowers its clock during a long game.
Static Power: Leakage Even When Idle
- P_static = I_static x V
- Example: I_static = 20 A, V = 1.1 V
- P_static = 20 x 1.1 = 22 W (used even when idle)
How big?
About 25% of total power, up to 50% in designs with very small transistors (Hennessy and Patterson).
Fix: Power gating turns off the supply to idle blocks. Like switching off the main switch, not just the fan.
Leakage (current that flows through a transistor even when it is off).
Module 4: The End of Dennard Scaling
Why clock speeds stopped rising and every chip became multicore
Dennard Scaling and the Power Wall
Until about 2004
After about 2005
- Dennard scaling: smaller transistors use lower voltage, so power per mm² stays constant.
- Clock rates rose about 40% per year.
- One faster core every generation.
- Voltage stuck near 1 V (leakage).
- Power density rose: the power wall.
- Clock rates stay around 3-5 GHz.
- Extra transistors become more cores (Unit 4).
Result: Performance now comes from parallelism and energy efficiency, not from a faster clock.
How Designers Save Energy Today
Clock gating
DVFS
Stop the clock to idle units, e.g., the FP unit when not used.
Lower V and f when work is light; raise them for games.
Sleep modes
Turbo boost
Phones are idle most of the time: deep sleep for DRAM and flash.
Run one core faster for a short time when others are idle.
Dark silicon: Not all transistors can be powered at once, so chips add special blocks (video, AI) that switch on only when needed.
Computer Usage Trends
Phones (PMD)
Cloud (WSC)
New workloads
For most people in Pakistan the first computer is a phone. Limited by the battery.
Apps like JazzCash and Easypaisa run on servers in data centers. Limited by electricity and cooling.
AI, video, payments need parallel and special-purpose hardware (GPUs, AI blocks).
So: At both ends of computing, energy is the main limit on design.
Class Activity: Four Quick Numericals
5 min · pairs
Solve in pairs with a calculator. Show the formula you use.
A chip has 20 billion transistors in 2026. If the count doubles every 2 years, how many in 2032?
Voltage falls from 1.2 V to 1.0 V, frequency unchanged. New dynamic power as a % of old?
Frequency falls by 25%, voltage unchanged. What happens to power, and to energy per task?
Latency improves 5x. By the rule of thumb, how much should bandwidth improve?
Common Mistakes
Wrong idea
Correct idea
✗ Moore's law says computers get twice as fast every two years.
✓ It says the number of transistors doubles about every two years.
✗ Power and energy are the same thing.
✓ Power is the rate (watts); energy is the total (joules, kWh).
✗ Lowering the clock always saves battery.
✓ Lower f alone saves power, not energy per task; lower V saves energy.
✗ If bandwidth doubles, latency halves.
✓ They are different: bandwidth grows much faster than latency improves.
Recap Quiz
Answer on paper in 3 minutes, then we check together.
Q1
Moore's law describes the growth of which quantity?
Q2
Write the formula for dynamic power.
Q3
Voltage falls by 10% and frequency stays the same. By what factor does dynamic energy change?
Q4
Why did clock rates stop rising around 2004-2005?
Q5
Give one technique a phone uses to save energy when it is idle.
Homework and Next Lecture
- Read Hennessy and Patterson, Chapter 1, sections 1.4 and 1.5.
- Numerical: a processor runs at 1.1 V and 2.5 GHz. DVFS lowers it to 0.9 V and 1.5 GHz. Find the new dynamic power and dynamic energy per task as a fraction of the old.
- Your phone: find its processor's manufacturing process (e.g., 4 nm) and maximum clock rate. Does it have big and small cores? Why?
- Think: why do data centers care about static (idle) power more than phones do?
NEXT LECTURE
Lecture 3: Cost Analysis and Trends in Cost
Learning curve and volume, the cost of an integrated circuit (wafer, die, yield), cost vs price, and dependability (MTTF, MTTR, availability).