Nexus Learn

Lecture 2: Technology Trends and Computer Usage

Lecture Notes CS-704 Advanced Computer Architecture 10 min read Updated 8 Nov 2026

↓ Download notes (Word)

On this page
  1. Learning Outcomes
  2. §1 Technology Trends
  3. §2 Bandwidth and Latency
  4. §3 Power and Energy
  5. §4 The End of Dennard Scaling
  6. §5 Common Mistakes
  7. Word Bank
  8. Self-Check Questions
  9. Further Reading

Learning Outcomes

SLO

After this lecture you can…

CLO

Bloom

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

§1.1 Moore's law

Slide L02-S06

Gordon Moore (co-founder of Intel) observed that the number of transistors on a chip doubles about every two years. It is a statement about the number of transistors, not about speed. Designers then decide how to use the extra transistors: bigger caches, more cores, wider pipelines, GPUs or AI blocks.

Worked example: checking Moore's law

Intel 4004 (1971): about 2,300 transistors
Apple M1 (2020): about 16 billion transistors

Growth = 16 x 10^9 / 2,300 = 6.96 million = 2^22.7 (22.7 doublings)
Years per doubling = (2020 - 1971) / 22.7 = 49 / 22.7 = 2.16 years

The result (about 2 years per doubling) matches Moore's prediction well. In recent years the pace has slowed, and chips also grow by stacking layers and joining several smaller dies (chiplets).

§1.2 How fast each technology improves

Slide L02-S07

Technology

Capacity growth

Doubling time

Where you see it

IC logic (transistors per chip)

40-55% per year

18-24 months

CPU and GPU chips

DRAM (main memory) per chip

25-40% per year

2-3 years (slowing)

8-16 GB RAM in laptops

Flash (SSD, phone storage)

50-60% per year

about 2 years

256 GB phone storage

Magnetic disk density

about 40% per year

about 2 years

data-center storage

Cost per bit: flash is about 15-20 times cheaper than DRAM, and disk is about 15-25 times cheaper than flash. Source: Hennessy and Patterson (2012), §1.4; newer editions report slower DRAM growth.

Design lesson: a new processor takes 3-5 years from first idea to the shop, so it must be designed for the technology that will exist when it ships, not the technology of today.

§1.3 Faster transistors, slower wires

Slide L02-S08

Feature size is the smallest width of a transistor or wire on a chip. It fell from about 10 micrometres (µm) in 1971 to a few nanometres (nm) today (1 nm = one billionth of a metre). Smaller transistors switch faster and use less energy per switch. Wires, however, become thinner and their resistance rises, so the delay of a signal crossing the chip improves much less. Wire delay is now a major limit on clock rate and chip layout. Note: process names such as '3 nm' are generation names, not an exact measured width.

Example: A wider GT Road lets more cars pass at the same time (more transistors), but the distance from Faisalabad to Lahore stays the same (wire delay).

§2 Bandwidth and Latency

§2.1 Two different questions

Slide L02-S10

Bandwidth (throughput)

Latency (response time)

Question

How much work or data per second?

How long does one task take?

Units

MB/s, GB/s, transactions per second

ns, ms, seconds

Motorway example

More lanes → more cars per hour

Travel time Lahore to Islamabad stays the same

Transport example

A Daewoo bus (50 seats) carries more people per trip

A car (4 seats) may reach sooner

§2.2 Why bandwidth improves faster

Slide L02-S11

Hennessy and Patterson's rule of thumb: bandwidth grows by at least the square of the improvement in latency.

bandwidth gain >= (latency gain)^2

Example: latency improves 4 times => bandwidth improves at least 4^2 = 16 times

Reason: bandwidth can be bought with parallelism (wider buses, more memory banks, more lanes), and extra transistors make parallelism cheap. Latency is limited by physics: distance, signal speed and the time to charge wires or rotate a disk. Because latency improves slowly, designers hide it with caches, prefetching and pipelining (Units 3 and 4).

§3 Power and Energy

§3.1 Power vs energy

Slide L02-S13

Power is the rate of using energy, measured in watts (1 W = 1 joule per second). Energy is the total amount, measured in joules or kilowatt-hours (kWh). On a WAPDA/FESCO electricity bill, one 'unit' is 1 kWh: a 1,000 W iron used for one hour uses 1 unit; a 100 W fan needs 10 hours to use the same unit.

Designer's concern

What it decides

Peak (maximum) power

The power supply must deliver it, or the voltage drops and the chip may fail.

Thermal design power (TDP)

The sustained power the cooling (heat sink, fan) must remove.

Energy per task

Battery life and electricity cost; the best metric for comparing two processors.

§3.2 Dynamic energy and dynamic power

Slide L02-S14

In CMOS chips most energy is used when transistors switch between 0 and 1.

Energy (one transition 0->1 or 1->0) = 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 (depends on transistors connected and wire length)
V = supply voltage f = switching frequency

  • Voltage: it is squared, so lowering voltage is the strongest way to save energy and power.
  • Frequency: changes power (energy per second) but not the energy per operation.

Example: C = 20 nF of switched capacitance, V = 1.0 V, f = 3 GHz: P = 0.5 × 20×10⁻⁹ × 1.0² × 3×10⁹ = 30 W.

§3.3 Worked example 1: lowering voltage by 15%

Slide L02-S15

New V = 0.85 x V_old and new f = 0.85 x f_old
Energy ratio = (0.85)^2 = 0.72 -> 28% less energy per operation
Power ratio = (0.85)^2 x 0.85 = 0.61 -> 39% less power

§3.4 Worked example 2: DVFS in your phone

Slide L02-S16

DVFS (dynamic voltage and frequency scaling) means the chip lowers its voltage and clock when work is light and raises them for heavy work. Thermal throttling is the same idea forced by heat: a phone that gets hot during a long game lowers its clock.

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 (the 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, because the task takes longer. Energy per task falls only when the voltage falls.

§3.5 Static power

Slide L02-S17

Static power is used even when transistors are not switching, because some current leaks through them (leakage). It grows with the number of transistors.

P_static = I_static x V
Example: I_static = 20 A, V = 1.1 V -> P_static = 22 W (even when idle)

Hennessy and Patterson report leakage as about 25% of total power, up to 50% in designs with very small transistors. Power gating (switching off the supply to idle blocks) reduces it.

§4 The End of Dennard Scaling

§4.1 Dennard scaling and the power wall

Slide L02-S19

Dennard scaling (1974) said that as transistors shrink, their voltage and current shrink too, so the power per square millimetre stays the same. This let clock rates rise by about 40% per year until around 2003. Around 2004-2005 the voltage could not be lowered much below about 1 V without too much leakage, so power density rose and chips became too hot to cool cheaply: the power wall. Intel cancelled a planned faster single-core processor in 2004 and moved to multicore. Since then clock rates have stayed at roughly 3-5 GHz, and extra transistors are used for more cores, larger caches and special-purpose units.

§4.2 How designers save energy today

Slide L02-S20

Technique

Idea

Example

Clock gating

Stop the clock to idle units

Floating-point unit switched off when only integer work runs

DVFS

Lower V and f when work is light

Phone slows down while you read messages

Low-power modes

Design for the typical (idle) case

DRAM and flash deep sleep when the screen is off

Turbo / overclocking

Run one core faster for a short time

Intel Turbo Boost while other cores are idle

Race-to-halt

Finish fast, then sleep

Burst at full speed, then deep sleep

Power gating

Cut supply to idle blocks

Reduces static (leakage) power

Dark silicon: on modern chips not all transistors can be powered at the same time within the power limit, so designers add special-purpose blocks (video encoders, AI engines) that are switched on only when needed.

§4.3 Computer usage trends

Slide L02-S21

Computer usage has moved to two ends: personal mobile devices (for most people in Pakistan the first computer is a smartphone) and cloud data centers (apps such as JazzCash and Easypaisa run on servers). Phones are limited by the battery and data centers by electricity and cooling costs. New workloads such as AI, video and digital payments need parallel and special-purpose hardware. At both ends energy efficiency is now a first-class design goal.

§5 Common Mistakes

Slide L02-S23

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 but not energy per task; lower V saves energy.

If bandwidth doubles, latency halves.

They are different quantities; bandwidth improves much faster than latency.

Word Bank

Word

Simple meaning

Real-world example

Bandwidth / throughput

Amount of work or data per unit time

Lanes on the motorway

Latency / response time

Time for one task from start to finish

Travel time from Lahore to Islamabad

Capacitive load (C)

How much electric charge must move to switch

Size of a water tank that must be filled

Dennard scaling

Smaller transistors need less voltage, so power density stays constant

Old rule that ended around 2005

DVFS

Changing voltage and frequency with the workload

Phone slows down when idle

Feature size

Smallest width of a transistor or wire

'4 nm' process in a new phone

Leakage

Current that flows even when a transistor is off

Charger plugged in with no phone

Moore's law

Transistor count doubles about every two years

2,300 (1971) to 16 billion (2020)

Power wall

Chips cannot use more power without overheating

Why clock speeds stopped rising

TDP

Thermal design power: sustained power the cooling must remove

65 W desktop CPU needs a heat sink and fan

Self-Check Questions

  • State Moore's law. Is it about speed or transistor count? (§1.1)
  • A chip has 20 billion transistors in 2026. Using doubling every two years, estimate the count in 2032. (§1.1)
  • Which technology improves fastest in capacity: DRAM, flash or disk? (§1.2)
  • Why does wire delay improve less than transistor speed? (§1.3)
  • Define bandwidth and latency and give one example of each. (§2.1)
  • Latency improves 5 times. How much should bandwidth improve by the rule of thumb? (§2.2)
  • What is the difference between power and energy? Which one is on the electricity bill? (§3.1)
  • Voltage falls from 1.2 V to 1.0 V with frequency unchanged. Find the new dynamic power as a fraction of the old. (§3.2)
  • Frequency falls by 25% with voltage unchanged. What happens to power and to energy per task? (§3.4)
  • A chip leaks 15 A at 0.9 V. Find its static power. (§3.5)
  • Why did clock rates stop rising around 2004-2005? (§4.1)
  • Explain DVFS, clock gating and turbo boost with one example each. (§4.2)

Answers: 160 billion; flash; at least 25 times; (1.0/1.2)² = 0.694; power 0.75, energy per task unchanged; 15 × 0.9 = 13.5 W.

Further Reading

  • Hennessy and Patterson (2012): Chapter 1, sections 1.4 (trends in technology) and 1.5 (trends in power and energy).

← All CS-704 material