Chapter 1: Fundamentals of Computer Design

1
Chapter 1 Fundamentals of Computer Design

• What is computer architecture?
• Why study computer architecture?
• Performance
• What is performance latency, throughput
• The performance equation
• Amdahl's law
• Measuring performance
• Cost

2
What is Computer Architecture?

• Instruction set architecture
• Instructions visible to programmer
• e.g., SPARC V8 vs. V9, Intel IA32 vs. IA64
• Organization
• Highlevel aspects of the system
• e.g., how many functional units, size of the
  cache, pipeline organization
• e.g., Ultra II vs. Ultra III
• e.g., Pentium III vs. Pentium 4
• Implementation or hardware
• Logic design
• e.g., 1.8 GHz vs. 2.4 GHz Pentium 4

3
Goals of the Computer Architect
4
Goals of the Computer Architect

• Depends on type of computer
• Supercomputer
• Server
• Desktop
• Embedded

5
Goals of the Computer Architect

• Functional goals
• Meet application area demands
• Compatability with previous systems
• Standards (e.g., IEEE floating point)
• Last through trends
• Performance
• Cost
• Power
• Energy
• Dependability
• Need to be familiar with design alternatives and
  criteria for selecting among them

6
Historical Trends

• Figure 1.1
• System performance is quadrupling every three
  years
• But this is not a law of nature!

7
Why Study Computer Architecture?
8
Why Study Computer Architecture?

• Technology changes fast and on different curves

9
Why Study Computer Architecture?

• Technology changes fast and on different curves
• Applications change
• From scientific to personal computing, databases,
  graphics, multimedia, communications
• Compiler / hardware boundary shifts
• New languages (e.g., shift from assembly to
  high-level languages)
• For many apps, the exponential curve is not
  enough!
• Parallel computing

10
Relationship to Prerequisites

• Prerequisite
• How to design a uniprocessor?
• This course
• How to design a uniprocessor WELL?
• Focus on performance
• Emphasis on Quantitative vs. Qualitative
• Common parallel architectures
• Be sure to check the handout for details on the
  prerequisites

11
What is Performance?

• Two Metrics
• Latency (or response time or execution time)
• Throughput (or bandwidth)

12
What is Performance?

• Two Metrics
• Latency (or response time or execution time)
• Time from start to finish of a task
• Throughput (or bandwidth)

13
What is Performance?

• Two Metrics
• Latency (or response time or execution time)
• Time from start to finish of a task
• Throughput (or bandwidth)
• Rate of task completion
• Rate of task initiation
• 1 / (time between task completions)

14
Performance (Cont.)

• Definition X is n faster than Y if
• Example X 1 minute, Y 2 minutes
• X is 100 faster than Y
• Example Automobile assembly line starts one car
  per hour and holds 20 cars
• Latency 20 hours
• Throughput one car per hour
• Throughput gt 1/Latency due to overlap

Execution TimeY Execution TimeX
n 100
1
15
Key Performance Equation
instructions cycles time
program instruction cycle
CPUtime X
X

• Instructions per program (path length)
• ISA and compiler
• Cycles per instruction (CPI)
• ISA and organization (e.g., cache misses)
• Time per cycle (clock time, cycle time)
• Organization and hardware

16
Amdahl's Law

• (Or why the common case matters most)
• Let
• Consider an enhancement x that speeds up fraction
  fx of a task by Sx
• Amdahls law gives

new rate old latency old rate new
latency
Speedup
old latency new latency
Speedupoverall
(1 - fx) (fx) ? old latency (1 - fx) ? old
latency fx /Sx ? old latency

1 (1 - fx) fx /Sx
Speedupoverall
17
Amdahl's Law, cont.

• Example fx 95 and Sx 1.10
• Example fx 5 and Sx 10
• Example fx 5 and Sx ?

1 (1 - 0.95) (0.95/1.10)
Speedupoverall
1.094
1 (1 - 0.05) (0.05/10)
Speedupoverall
1.047
1 (1 - 0.05) (0.05/?)
Speedupoverall
1.052
18
Amdahl's Law Corollary

• Since Sx ? ? implies Example
• For all real speedups
• Or make the common case fast
• An application?

1 (1 - fx) (fx /?)
Speedupoverall
1 1 - fx
Speedupoverall lt
19
Measuring Performance

• MIPS, MFLOPS don't mean much
• Benchmarks
• Real programs
• Representative of real workload
• Only way to characterize performance
• SPEC89 ? SPEC92 ? SPEC95 ? SPEC CPU2000
• TPC
• Kernels
• Representative'' program fragments
• Often not representative of full applications
• EEMBC for embedded systems
• Toy benchmarks and synthetic benchmarks
• Don't mean much

20
Cost

• Cost is very important in most real designs
• But usually hard to quantify for the architect
• Costs change over time
• Learning curve lowers manufacturing costs
• Technology improvements lower costs
• E.g., DRAM generation price falls by 10 to 30x
  over lifetime
• Figures 1.5 and 1.6
• Focus on IC costs bigger price variable

21
Integrated Circuit Cost
Cost of Die Cost of Testing Cost of
Packaging Final Test Yield
Cost of IC
Cost of Wafer Dies per Wafer ? Die Yield
Cost of Die
? ? (Wafer Diameter/2)2 Die Area
Dies per Wafer (
)
(Correction factor for Edge Effects)
Defects per unit area ? Die Area ?
Die Yield Wafer Yield ?1
- ?
Cost of Die f(Die area)5, assuming ? 4
22
Cost Breakdown

• Component Cost
• Microprocessor, SRAM, DRAM
• Disk
• Power supplies and packaging
• Direct Costs
• Manufacturing (labor)
• Warranty
• Indirect Costs or Gross Margin
• Research and Development
• Sales and Marketing
• Profits and Taxes

23
Price

• Only loosely related to cost
• Start with all component costs
• Add 10 to 30 for direct costs
• Add 10 to 80 gross margin (indirect costs)
• AVERAGE SELLING PRICE
• Add discounts and dealer profit
• LIST PRICE
• Note
• 1 increase in component can imply
• 1.21 to 2.34 increase in ASP
• 1.57 to 3.04 increase in list price (if
  discount 30 ASP)
• Component cost 33 64 of list price
• RD is often 4 to 12 of list price