COMPUTER ORGANIZATION AND DESIGN

RISC-V
Edition
The Hardware/Software Interface

Chapter 1
Computer Abstractions
and Technology
The Computer Revolution

§1.1 Introduction
 Progress in computer technology
 Underpinned by Moore’s Law
 Makes novel applications feasible
 Computers in automobiles
 Cell phones
 Human genome project
 World Wide Web
 Search Engines
 Computers are pervasive

2
Classes of Computers
 Personal computers
 General purpose, variety of software
 Subject to cost/performance tradeoff

 Server computers
 Network based
 High capacity, performance, reliability
 Range from small servers to building sized

3
Classes of Computers
 Supercomputers
 High-end scientific and engineering
calculations
 Highest capability but represent a small
fraction of the overall computer market

 Embedded computers
 Hidden as components of systems
 Stringent power/performance/cost constraints

4
The PostPC Era

5
The PostPC Era
 Personal Mobile Device (PMD)
 Battery operated
 Connects to the Internet
 Hundreds of dollars
 Smart phones, tablets, electronic glasses
 Cloud computing
 Warehouse Scale Computers (WSC)
 Software as a Service (SaaS)
 Portion of software run on a PMD and a
portion run in the Cloud
 Amazon and Google

6
What You Will Learn
 How programs are translated into the
machine language
 And how the hardware executes them
 The hardware/software interface
 What determines program performance
 And how it can be improved
 How hardware designers improve
performance
 What is parallel processing

7
Understanding Performance
 Algorithm
 Determines number of operations executed
 Programming language, compiler, architecture
 Determine number of machine instructions executed
per operation
 Processor and memory system
 Determine how fast instructions are executed
 I/O system (including OS)
 Determines how fast I/O operations are executed

8
Eight Great Ideas

§1.2 Eight Great Ideas in Computer Architecture

 Design for Moore’s Law
 Use abstraction to simplify design
 Make the common case fast
 Performance via parallelism
 Performance via pipelining
 Performance via prediction
 Hierarchy of memories
 Dependability via redundancy

9
Below Your Program

§1.3 Below Your Program

 Application software
 Written in high-level language
 System software
 Compiler: translates HLL code to
machine code
 Operating System: service code
 Handling input/output
 Managing memory and storage
 Scheduling tasks & sharing resources
 Hardware
 Processor, memory, I/O controllers

10
Levels of Program Code
 High-level language
 Level of abstraction closer
to problem domain
 Provides for productivity
and portability
 Assembly language
 Textual representation of
instructions
 Hardware representation
 Binary digits (bits)
 Encoded instructions and
data

11
Components of a Computer

§1.4 Under the Covers

The BIG Picture  Same components for
all kinds of computer
 Desktop, server,
embedded
 Input/output includes
 User-interface devices
 Display, keyboard, mouse
 Storage devices
 Hard disk, CD/DVD, flash
 Network adapters
 For communicating with
other computers

12
Touchscreen
 PostPC device
 Supersedes keyboard
and mouse
 Resistive and
Capacitive types
 Most tablets, smart
phones use capacitive
 Capacitive allows
multiple touches
simultaneously

13
Through the Looking Glass
 LCD screen: picture elements (pixels)
 Mirrors content of frame buffer memory

14
Opening the Box
Capacitive multitouch LCD screen

3.8 V, 25 Watt-hour battery

Computer board

15
Inside the Processor (CPU)
 Datapath: performs operations on data
 Control: sequences datapath, memory, ...
 Cache memory
 Small fast SRAM memory for immediate
access to data

16
Inside the Processor
 Apple A5

17
Abstractions
The BIG Picture

 Abstraction helps us deal with complexity

 Hide lower-level detail
 Instruction set architecture (ISA)
 The hardware/software interface
 Application binary interface
 The ISA plus system software interface
 Implementation
 The details underlying and interface

18
A Safe Place for Data
 Volatile main memory
 Loses instructions and data when power off
 Non-volatile secondary memory
 Magnetic disk
 Flash memory
 Optical disk (CDROM, DVD)

19
Networks
 Communication, resource sharing,
nonlocal access
 Local area network (LAN): Ethernet
 Wide area network (WAN): the Internet
 Wireless network: WiFi, Bluetooth

20
Technology Trends

§1.5 Technologies for Building Processors and Memory

 Electronics
technology
continues to evolve
 Increased capacity
and performance
 Reduced cost
DRAM capacity

Year Technology Relative performance/cost

1951 Vacuum tube 1
1965 Transistor 35
1975 Integrated circuit (IC) 900
1995 Very large scale IC (VLSI) 2,400,000
2013 Ultra large scale IC 250,000,000,000

21
Semiconductor Technology
 Silicon: semiconductor
 Add materials to transform properties:
 Conductors
 Insulators
 Switch

22
Manufacturing ICs

 Yield: proportion of working dies per wafer

23
Intel Core i7 Wafer

 300mm wafer, 280 chips, 32nm technology

 Each chip is 20.7 x 10.5 mm
24
Integrated Circuit Cost

 Nonlinear relation to area and defect rate

 Wafer cost and area are fixed
 Defect rate determined by manufacturing process
 Die area determined by architecture and circuit design

25
Defining Performance

§1.6 Performance
 Which airplane has the best performance?

26
Response Time and Throughput
 Response time
 How long it takes to do a task
 Throughput
 Total work done per unit time
 e.g., tasks/transactions/… per hour
 How are response time and throughput affected
by
 Replacing the processor with a faster version?
 Adding more processors?
 We’ll focus on response time for now…

27
Relative Performance
 Define Performance = 1/Execution Time
 “X is n time faster than Y”

 Example: time taken to run a program

 10s on A, 15s on B
 Execution TimeB / Execution TimeA
= 15s / 10s = 1.5
 So A is 1.5 times faster than B

28
Measuring Execution Time
 Elapsed time
 Total response time, including all aspects
 Processing, I/O, OS overhead, idle time
 Determines system performance
 CPU time
 Time spent processing a given job
 Discounts I/O time, other jobs’ shares
 Comprises user CPU time and system CPU
time
 Different programs are affected differently by
CPU and system performance

29
CPU Clocking
 Operation of digital hardware governed by a
constant-rate clock
Clock period

Clock (cycles)

Data transfer
and computation
Update state

 Clock period: duration of a clock cycle

 e.g., 250ps = 0.25ns = 250×10–12s
 Clock frequency (rate): cycles per second
 e.g., 4.0GHz = 4000MHz = 4.0×109Hz

30
CPU Time

 Performance improved by
 Reducing number of clock cycles
 Increasing clock rate
 Hardware designer must often trade off clock
rate against cycle count

31
CPU Time Example
 Computer A: 2GHz clock, 10s CPU time
 Designing Computer B
 Aim for 6s CPU time
 Can do faster clock, but causes 1.2 × clock cycles
 How fast must Computer B clock be?

32
Instruction Count and CPI

 Instruction Count for a program

 Determined by program, ISA and compiler
 Average cycles per instruction
 Determined by CPU hardware
 If different instructions have different CPI
 Average CPI affected by instruction mix

33
CPI Example
 Computer A: Cycle Time = 250ps, CPI = 2.0
 Computer B: Cycle Time = 500ps, CPI = 1.2
 Same ISA
 Which is faster, and by how much?

A is faster…

…by this much

34
CPI in More Detail
 If different instruction classes take different
numbers of cycles

 Weighted average CPI

Relative frequency

35
CPI Example
 Alternative compiled code sequences using
instructions in classes A, B, C
Class A B C
CPI for class 1 2 3
IC in sequence 1 2 1 2
IC in sequence 2 4 1 1

 Sequence 1: IC = 5  Sequence 2: IC = 6
 Clock Cycles  Clock Cycles
= 2×1 + 1×2 + 2×3 = 4×1 + 1×2 + 1×3
= 10 =9
 Avg. CPI = 10/5 = 2.0  Avg. CPI = 9/6 = 1.5

36
Performance Summary
The BIG Picture

 Performance depends on
 Algorithm: affects IC, possibly CPI
 Programming language: affects IC, CPI
 Compiler: affects IC, CPI
 Instruction set architecture: affects IC, CPI, Tc

37
Power Trends

§1.7 The Power Wall

 In CMOS IC technology

×30 5V → 1V ×1000

38
Reducing Power
 Suppose a new CPU has
 85% of capacitive load of old CPU
 15% voltage and 15% frequency reduction

 The power wall

 We can’t reduce voltage further
 We can’t remove more heat
 How else can we improve performance?

39
Uniprocessor Performance

§1.8 The Sea Change: The Switch to Multiprocessors

Constrained by power, instruction-level parallelism,
memory latency

40
Multiprocessors
 Multicore microprocessors
 More than one processor per chip
 Requires explicitly parallel programming
 Compare with instruction level parallelism
 Hardware executes multiple instructions at once
 Hidden from the programmer
 Hard to do
 Programming for performance
 Load balancing
 Optimizing communication and synchronization

41
SPEC CPU Benchmark
 Programs used to measure performance
 Supposedly typical of actual workload
 Standard Performance Evaluation Corp (SPEC)
 Develops benchmarks for CPU, I/O, Web, …
 SPEC CPU2006
 Elapsed time to execute a selection of programs
 Negligible I/O, so focuses on CPU performance
 Normalize relative to reference machine
 Summarize as geometric mean of performance ratios
 CINT2006 (integer) and CFP2006 (floating-point)

42
CINT2006 for Intel Core i7 920

43
SPEC Power Benchmark
 Power consumption of server at different
workload levels
 Performance: ssj_ops/sec
 Power: Watts (Joules/sec)

44
SPECpower_ssj2008 for Xeon X5650

45
Pitfall: Amdahl’s Law

§1.10 Fallacies and Pitfalls

 Improving an aspect of a computer and
expecting a proportional improvement in
overall performance

 Example: multiply accounts for 80s/100s

 How much improvement in multiply performance to
get 5× overall?

 Can’t be done!

 Corollary: make the common case fast

46
Fallacy: Low Power at Idle
 Look back at i7 power benchmark
 At 100% load: 258W
 At 50% load: 170W (66%)
 At 10% load: 121W (47%)
 Google data center
 Mostly operates at 10% – 50% load
 At 100% load less than 1% of the time
 Consider designing processors to make
power proportional to load

47
Pitfall: MIPS as a Performance Metric
 MIPS: Millions of Instructions Per Second
 Doesn’t account for
 Differences in ISAs between computers
 Differences in complexity between instructions

 CPI varies between programs on a given CPU

48
Concluding Remarks

§1.11 Concluding Remarks

 Cost/performance is improving
 Due to underlying technology development
 Hierarchical layers of abstraction
 In both hardware and software
 Instruction set architecture
 The hardware/software interface
 Execution time: the best performance
measure
 Power is a limiting factor
 Use parallelism to improve performance

49