Nexus Learn

Lecture 3: Cost Analysis and Trends in Cost

Lecture Slides CS-704 Advanced Computer Architecture 6 min read Updated 8 Nov 2026

↓ Download slides (PowerPoint)

Recap of Lecture 2 and Homework Check

  • Moore's law: transistor count doubles about every two years.
  • Bandwidth grows at least as the square of latency improvement.
  • Dynamic power = 1/2 x C x V^2 x f; static power = I_static x V.
  • Lowering frequency alone saves power, not energy per task.

Homework check

  • 1.1 V, 2.5 GHz -> 0.9 V, 1.5 GHz.
  • New power and energy per task as a fraction of the old?
  • Why do data centers care about idle (static) power more than phones?

Today's Plan (50 minutes)

0-5

Recap and homework check Lecture 2

5-8

Plan and learning outcomes

8-14

Module 1: Trends in cost learning curve, volume, commodities

14-26

Module 2: Cost of a chip wafer, dies, yield, die cost

26-35

Module 3: Dependability MTTF, MTTR, availability, redundancy

35-43

Class activity four short numericals

43-50

Mistakes, quiz and homework

Learning Outcomes (SLOs)

By the end of this lecture you will be able to:

SLO 3.1

Explain how the learning curve, volume and commoditization lower the cost of computers.

→ CLO1 · C2 Understand

SLO 3.2

Calculate dies per wafer, die yield and the cost of a die for a given chip size.

→ CLO2 · C3 Apply

SLO 3.3

Calculate MTTF, failure rate and availability of a system and show how redundancy helps.

→ CLO2 · C3 Apply

  • 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: Trends in Cost

Why the same computer gets cheaper every year

The Learning Curve: Cost Falls Over Time

Learning curve

Real-world example

  • As a factory makes more chips, its yield (percentage of good chips) rises.
  • More good chips from the same wafer = lower cost per chip.
  • A new tandoor wastes many burnt rotis in its first week; after a month, almost none.
  • A phone launched at Rs 120,000 often sells for much less a year later.

Key word: Yield (percentage of manufactured items that work) is the single biggest factor in chip cost.

Volume and Commoditization

Volume

Commodities

  • More units = faster learning, cheaper buying, and design cost shared over more units.
  • Rule of thumb: cost falls about 10% per doubling of volume.
  • Many vendors sell nearly identical products: DRAM, flash, disks, keyboards.
  • Competition pushes price close to cost.

Example: Mobile chargers and USB drives in any electronics market: many sellers, same product, low price - a commodity.

Module 2: The Cost of a Chip

From a round silicon wafer to a packaged processor

From Wafer to Chip

Wafer

Dies

Test

Cut and package

Final test

Round silicon disc, 30 cm

Many chips made at once

Find bad dies

Good dies into cases

Ship working chips

Roti analogy

Cutting square pieces from a round roti wastes the edges. A burnt spot ruins whichever piece it falls on - and a bigger piece is more likely to contain one. That is why big chips cost much more.

The Cost Formulas

  • Cost of die = Cost of wafer / (Dies per wafer x Die yield)
  • Dies per wafer = pi x (d/2)^2 / A - pi x d / sqrt(2 x A)
  • (whole wafer) (edge waste)
  • Die yield = Wafer yield x 1 / (1 + D x A)^N
  • d = wafer diameter A = die area D = defects per cm^2
  • N = process-complexity factor (about 11.5-15.5)

Cost of IC = (cost of die + testing + packaging and final test) / final test yield. Source: Hennessy and Patterson (2012), §1.6.

Worked Example: Big Die vs Small Die

4x

the cost

for 2.25x the area

Lesson: A bigger die gets fewer dies per wafer AND lower yield, so cost grows faster than area.

Cost vs Price

Cost

Price

  • What it takes to make one unit: die, testing, packaging, assembly.
  • Falls with yield, volume and the learning curve.
  • What the customer pays: cost + R&D + marketing + profit (gross margin).
  • NRE (one-time design and mask cost) is shared over all units.

Example: A chip made only for a small market (a few thousand cameras) carries a huge share of NRE per unit; a mass-market chip does not.

Module 3: Dependability

How often does it fail, and how fast is it repaired?

Measuring Dependability

Example: A UPS that works 2,000 hours between failures and takes 48 hours to repair: availability = 2000 / 2048 = 97.7%.

Worked Example: MTTF of a Disk Subsystem

  • Failure rate = 10 x 1/1,000,000 (10 disks)
  • + 1/500,000 (controller)
  • + 1/200,000 (power supply)
  • + 1/200,000 (fan)
  • + 1/1,000,000 (cable)
  • = 23 / 1,000,000 per hour = 23,000 FIT
  • System MTTF = 1,000,000 / 23 = 43,478 hours (about 5 years)

43,478 h

system MTTF

less than any single part

Assumes failures are independent and failure rates are constant, so rates add.

Redundancy: Two Power Supplies

  • MTTF(pair) = MTTF^2 / (2 x MTTR)
  • MTTF = 200,000 h, MTTR = 24 h
  • MTTF(pair) = 200,000^2 / (2 x 24) = 833,333,333 h
  • = about 4,200 times one supply

Why it works

Both must fail at the same time - very unlikely if repair is quick.

Example: A house with a UPS and a generator for load shedding: power is lost only if both fail together.

Class Activity: Four Quick Numericals

7 min · pairs

Solve in pairs with a calculator. Show the formula you use.

Dies per wafer: wafer 20 cm, die area 1.0 cm^2.

Die yield: D = 0.05 per cm^2, die area 1.0 cm^2, N = 13.5 (wafer yield 100%).

A server has MTTF 50,000 h and MTTR 10 h. Find its availability.

Failure rates: 2 parts with MTTF 100,000 h, one 20,000 h, one 40,000 h. Find system MTTF.

Common Mistakes

Wrong idea

Correct idea

✗ A die twice as big costs twice as much.

✓ Cost grows faster than area: fewer dies per wafer and lower yield.

✗ Cost and price are the same.

✓ Price = cost + R&D, marketing and profit (gross margin).

✗ System MTTF = average of part MTTFs.

✓ Add the failure rates (1/MTTF), then take 1 / total.

✗ MTBF and MTTF are the same.

✓ MTBF = MTTF + MTTR.

Recap Quiz

Answer on paper in 3 minutes, then we check together.

Q1

What is yield, and why does it rise over time?

Q2

Write the formula for the cost of a die.

Q3

Why does a bigger die cost more than its area alone suggests?

Q4

A system has MTTF 9,990 h and MTTR 10 h. What is its availability?

Q5

Why does adding a second power supply make a system so much more reliable?

Homework and Next Lecture

  • Read Hennessy and Patterson, Chapter 1, sections 1.6 and 1.7.
  • Numerical: 30 cm wafer costing $6,000, die area 1.5 cm^2, D = 0.04 per cm^2, N = 13.5. Find dies per wafer, die yield and cost per die.
  • Dependability: a web server has 2 disks (MTTF 1,000,000 h each), a power supply (200,000 h) and a motherboard (300,000 h). Find system MTTF and FIT.

NEXT LECTURE

Lecture 4: Measuring and Reporting Performance

Execution time vs throughput, and the CPU performance equation: time = instruction count x CPI x clock cycle time.

← All CS-704 material