An Introduction to

Computer Architecture

Module 1

© 2019 Arm Limited

This Course
• Aims to introduce the key concepts and ideas in computer architecture
• Explores the design of modern microprocessors
• Examines important trends and current and future challenges

2 © 2019 Arm Limited

Module Syllabus
• What is computer architecture?
• Computer architecture arena and design goals
• Historical performance of computer architecture
• Future trends with multicore processors, systems on chip (SoCs), and beyond

3 © 2019 Arm Limited

Introduction
• The modern microprocessor is often the engine behind
how we communicate and work.
• It has helped to create our digital world where
communication, computation, and storage is almost free.
• It underpins many scientific breakthroughs and can help
Die photo of Arm Cortex-M3 Microcontroller
us make better use of the world’s limited resources with 16KB flash memory and 4KB RAM 1

Google Data Center2

1. By Zeptobars, CC BY 3.0
4 © 2019 Arm Limited 2. By Connie Zhou, CC BY-NC 4.0
Introduction
• The modern computer is less than 100 years old.
• The first electromechanical and valve-based machines
were produced in the 1930s and 1940s.
• Today’s machines are many orders of magnitude faster,
EDSAC replica (2018)1
lower power, more reliable, and cheaper.

BBC Micro Bit (2015) Arm Cortex-M02

1. EDSAC photo, CC BY-SA 4.0

5 © 2019 Arm Limited 2. Aruld, CC BY-SA 4.0
Levels of Abstraction
• Architecture
• A set of specifications that allows developers to write software and firmware
• These include the instruction set.
• Microarchitecture
• The logical organization of the inner structure of the computer
• Hardware or Implementation
• The realization or the physical structure, i.e., logic design and chip packaging

6 © 2019 Arm Limited

Architecture Specifications
• The architecture specifies a contract between
the hardware and software.
• Many different compatible processors may be
implemented, e.g., to meet different power
consumption, cost, area, and performance
goals.
• When a software is written to conform with
an architecture specification, it can be
portable.
IBM System/360
Source of photo: Erik Pitti, CC-BY-2.0

7 © 2019 Arm Limited

Computer Architecture
• Computer architecture is much more than the task of defining the instruction-set or
high-level architecture.
• The computer architect must contribute to, and understand, all levels of design in order
to deliver the most appropriate design for a particular application and target market.

8 © 2019 Arm Limited

Computer Architecture
• Computer architecture is concerned with how best to exploit fabrication technology to
meet marketplace demands.
• e.g., how best might we use five billion transistors and a power budget of two watts to design the
chip at the heart of a mobile phone?
• Computer architecture builds on a few simple concepts, but is challenging as we must
constantly seek new solutions.
• What constitutes the “best” design changes over time and depending on our use-case.
It involves considering many different trade-offs.

9 © 2019 Arm Limited

Computer Architecture
• Each level of design imposes different
requirements and constraints, which change
over time. Markets
• History and economics: there is commercial Applications
pressure to evolve in a way that minimizes Operating Systems
disruption and possible costs to the ecosystem Programming languages and
(e.g., software). compilers
• There is also a need to look forward and not Architecture
design for yesterday’s technology and Microarchitecture
workloads! Hardware
• Design decisions should be carefully justified Fabrication Technology
through experimentation.

10 © 2019 Arm Limited

The Computer Architecture Arena

Computer
architecture
Application
characteristics
Markets

New
applications

Technology

Source: “Early 21st Century Processors,” S. Vajapeyam and M. Valero, IEEE Computer, April 2004
11 © 2019 Arm Limited
Design Goals I
• Functional – hard to correct (unlike software). Verification is perhaps the highest single
cost in the design process. We also need to test our chips once they have been
manufactured, again this can be a costly process and requires careful thought at the
design stage
• Performance – what does this mean? No single best answer, e.g., sports car vs. off-road
4x4 vehicle – performance will always depend on the “workload”
• Power – a first-order design constraint for most designs today. Power limits the
performance of most systems.

12 © 2019 Arm Limited

Design Goals II
• Security – e.g., the ability to control access to sensitive data or prevent carefully crafted
malicious inputs from hijacking control of the processor
• Cost – design cost (complexity), die costs (i.e., the size or area of our chip), packaging,
etc.
• Reliability – do we need to try to detect and/or tolerate faults during operation?

13 © 2019 Arm Limited

Markets and Features
Each target market will require a different trade-off in terms of power consumption, cost,
area, performance, security, reliability, etc.
Here are some example processor classes from Arm:
• Cortex-A: high-performance application processors, e.g., for mobile phones
• Cortex-R: deterministic real-time performance, fault detection, and tolerance.
• Cortex-M: energy-efficient embedded devices (“microcontroller” class cores)
• Neoverse: scalable networks of processors on a single chip
• e.g., 8, 16, 64, or 128 cores. Used in datacenters, edge servers, and storage

14 © 2019 Arm Limited

 Camera
The Smartphone CORTEX-M Sensor hub
Touchscreen & CORTEX-M Power management
sensor hub
CORTEX-M
CORTEX-A
• A single CORTEX-M
smartphone will Flash controller
Apps processor
contain many CORTEX-M
CORTEX-A
different CORTEX-M
processor cores.
GPS
2G/3G/4G/5G CORTEX-M
• Why not use a CORTEX-A
single processor? CORTEX-R
CORTEX-M Bluetooth
CORTEX-M
Wi-Fi
CORTEX-R
CORTEX-M
15 © 2019 Arm Limited
 Historical Performance

16 © 2019 Arm Limited

Historical Performance Gains
• By 1985, it was possible to integrate a complete microprocessor onto a single die or
“chip.”
• As fabrication technology improved, and transistors got smaller, the performance of a
single core improved quickly.
• Performance improved at the rate of 52% per year for nearly 20 years (measured using
SPEC benchmark data).
• Note: the data are for desktop/server processors

17 © 2019 Arm Limited

Historical Performance Gains
Clock period
• Clock frequency improved quickly
between 1985 and 2002:

Clock Frequency (MHz)

• ~10x from faster transistors, and
• ~10x from pipelining and circuit-level
advances.
• So overall, ~100X of the total 800X gains
came from reduced clock periods.

Year
A. Danowitz, K. Kelley, J. Mao, J. P. Stevenson, and M. Horowitz.
Clock Frequency, Stanford CPU DB. Accessed on  Nov. 5, 2019.
[Online]. Available:
http://cpudb.stanford.edu/visualize/clock_frequency 
18 © 2019 Arm Limited
Historical Performance Gains
• From 1985 to 2002, performance improved by ~800 times.
• Over time, technology scaling provided much greater numbers of faster and lower
power transistors.
• The “iron law” of processor performance:

Time = instructions executed x clocks per instruction (CPI) x clock period

• Clocks per instruction (CPI)

• We will also refer to Instructions Per Cycle (IPC), i.e., 1/CPI.

19 © 2019 Arm Limited

Clocks Per Instruction (CPI)
• Early machines were limited by transistor count. As a result, they often required
multiple clock cycles to execute each instruction (CPI >> 1).
• As transistor budgets improved, we could aim to get closer to a CPI of 1.
• This is easy if we don’t care at all about clock frequency.
• Designing a high-frequency design with a good CPI is much harder. We need to keep our
high-performance processor busy and avoid it stalling, which would increase our CPI.
This requires many different techniques and costs transistors (area) and power.

20 © 2019 Arm Limited

Clocks Per Instruction (CPI)
• Eventually, the industry was also able to fetch and execute multiple instructions per
clock cycle. This reduced CPI to below 1.
• When we fetch and execute multiple instructions together, we often refer to
Instructions Per Cycle (IPC), which is 1/CPI.
• For instructions to be executed at the same time, they must be independent.
• Again, growing transistor budgets were exploited to help find and exploit this
Instruction-Level Parallelism (ILP).

21 © 2019 Arm Limited

 What Is Pipelining?
Engine
Chassis Paint
Time Time
A A
B B
C C

Order of manufacturing Order of

(Car A, B, and then C) manufacturing
22 © 2019 Arm Limited
IPC and Instruction Count
• Of the 800x improvement in
performance (1985-2002), ~100x is SpecInt2000 per MHz
from clock frequency improvements. 0

• The remaining gains (~8x) were from 0

a reduction in instruction count, 0

better compiler optimizations, and 0

improvements in IPC. 0

The graph to the right shows these improvements. It plots 0

0 5 10 15 20 25
performance (SpecInt2000 benchmark performance per
MHz for Intel processors against time).
Year

23 © 2019 Arm Limited

A Shorter Critical Path
• We can also try to reduce the number of
gates on our critical path.

Critical path length (FO4 delays)

• This can be done by inserting additional
registers to break complex logic into
different “pipeline” stages.
• Advances were also made that improved
circuit-level design techniques.
• The length of our critical paths reduced by
~10x (1985-2002).

Year
A. Danowitz, K. Kelley, J. Mao, J. P. Stevenson, and M.
Horowitz. Stanford CPU DB. Accessed on  Nov. 5, 2019.
24 © 2019 Arm Limited
[Online]. Available: http://cpudb.stanford.edu 
Moore’s Law
• Moore’s Law predicts that the number of
transistors we can integrate onto a chip, for
the same cost, doubles every 2 years.

Gordon Moore and Robert Noyce at Intel in 1970

Source: IntelFreePress, CC BY-SA-2.0
25 © 2019 Arm Limited
Moore’s Law
• Processor transistor budgets grew
quickly as microarchitectures became
more complex.
• 1985 – Intel 386
275K transistors, die size = 43 mm2
• 2002 – Intel Pentium 4
42M transistors, die size = 217 mm2

26 © 2019 Arm Limited

Source: Wgsimon, Wikipedia, CC BY-SA 3.0
Better Performance and Lower Power

27 © 2019 Arm Limited

 Multicore Processors,
Systems on Chip (SoC), and Beyond

28 © 2019 Arm Limited

Slowing Single-core Performance Gains
To summarize, sustaining single core performance gains became difficult due to:
• The limits of pipelining
• The limits of Instruction-Level Parallelism (ILP)
• Power consumption
• The performance of on-chip wires

As a result performance gains slowed from 52% to 21% per year for the highest
performance processors.

29 © 2019 Arm Limited

Multicore Processors
• Eventually, it made sense to shift from
single-core to multicore designs.
• From ~2005, multicore designs became
mainstream.
• The number of cores on a single chip
increased over time.
• Clock frequencies increased more slowly.
• Individual cores were designed to be as
power efficient as possible.

e.g., 4 x Arm Cortex-A72 processors,

each with their own L1 caches and a
shared L2 cache
30 © 2019 Arm Limited
Multicore Processors
Exploiting multiple cores comes with its own set of challenges and limitations:
• Power consumption may still limit performance.
• We need to write scalable and correct parallel programs to exploit them.
• We might not be able to find enough parallel threads to take advantage of our cores.
• On-chip and off-chip communication will limit performance gains.
• Off-chip bandwidth is limited and may throttle our many cores.
• Cores also need to communicate to maintain a coherent view of memory.

31 © 2019 Arm Limited

Specialization
• Today, we often need to look beyond general-purpose programmable processors to
meet our design goals.
• We trade flexibility for efficiency.
• We remove the ability to run all programs and design for a narrow workload, perhaps
even a single algorithm.
• These “accelerators” can be 10-1000x better than a general-purpose solution in terms
of power and performance.

32 © 2019 Arm Limited

Specialization

Graphics Processing Unit Neural Processor Unit (NPU)

(GPU)
33 © 2019 Arm Limited
Today’s SoC Designs
• A modern mobile phone SoC (2019)
may contain more than 7 billion mem interface mem interface
transistors.
• It will integrate: L3 cache
• Multiple processor cores memory Neural
• Processor
A GPU
Unit
• A large number of specialized accelerators
GPU 4 “big” (NPU)
• Large amounts of on-chip memory cores
• High bandwidth interfaces to off-chip
memory
4 “small” Other
cores accelerators

mem interface mem interface

A high-level block diagram of a

34 © 2019 Arm Limited
mobile phone SoC
Trends in Computer Architecture

Early computers Gains from bit-level parallelism

Time
Pipelining and superscalar issue + Instruction-level parallelism
Multicore/GPUs + Thread-level parallelism/data-level
parallelism
Greater integration (large SoCs),
heterogeneity, and specialization + Accelerator-level parallelism

Note: Memory hierarchy developments have also been significant. The

memory hierarchy typically consumes a large fraction of the transistor
budget.
35 © 2019 Arm Limited
The Future – The End of Moore’s Law?
• The end of Moore’s Law has been predicted many times.
• Scaling has perhaps slowed in recent years, but transistor density continues to improve.
• Eventually, 2D scaling will have to slow down.
• We are ultimately limited by the size of atoms!
• Where next?
• Going 3D - Future designs may take advantage of multiple layers of transistors on a single chip.
– Note: the gains are linear rather than exponential.
• Better packaging and integration technologies (e.g., chip stacking)
• New types of memory
• New materials and devices

36 © 2019 Arm Limited

 Backup Slides

37 © 2019 Arm Limited

From Sensors and Smartphones to Servers
1 laptop or server class
An area optimized Mid-range 64-bit processor (e.g., A76 core with
microcontroller core processor (e.g., Arm 512KB of L2 cache)
(e.g., Arm Cortex-M0) Cortex-A55). For
1X smartphones, TVs,
network infrastructure,
1 square represents …
the area of this core 1380X

130X
High-performance
32-bit core
(e.g., Arm Cortex-M7) High-performance
13X
Used in automotive, processor (e.g., Arm
sensor hub, and other Cortex-A73). For
embedded applications. mobile and consumer
devices.
38 © 2019 Arm Limited 520X
Technology Scaling: Faster Transistors
• From 1985 to 2002, we saw ~7 new process
generations.
• Scaling provides smaller and faster
transistors. Performance improves ~1.4x

Feature Size (um)

per generation, so for 7 generations, we
have ~10x faster logic gates.

Year
A. Danowitz, K. Kelley, J. Mao, J. P. Stevenson, and M. Horowitz.
Stanford CPU DB. Accessed on  Nov. 5, 2019. [Online]. Available:
http://cpudb.stanford.edu 

39 © 2019 Arm Limited

Instruction Count
• Increased datapath width (e.g., 16-bit to 32-bit to 64-bit)
• Larger register files (fewer load/store instructions)
• More complex instructions?
• SIMD instructions

40 © 2019 Arm Limited

Limits to Single Core Performance
• Limits to pipelining
• Cost of interruptions grow, e.g., impact of cache misses and mispredicted branches.
• Ultimately, some components are difficult or expensive to pipeline.
• There are also practical limits to distributing very high-frequency clocks, registers represent a finite
delay, and we may struggle to balance logic between pipeline stages.
• Limits of Instruction-Level Parallelism (ILP)
• Large amounts of ILP are very difficult to discover and exploit efficiently.
• Our returns on investment quickly diminish, i.e., we must use more power and more transistors to
expose and exploit ever smaller amounts of ILP.

41 © 2019 Arm Limited

Limits to Single Core Performance
• Power consumption
• Historical performance gains have been
impressive, but power consumption also
grew very quickly during the 1980s and
1990s.
• This happened even with improvements in
fabrication technology and reductions in
supply voltage.
• Power quickly became, and remains, a first-
order design constraint for all significant
markets.

Year
Figure source: Original data collected and plotted by M. Horowitz, F. Labonte, O. Shacham, K.
Olukotun, L. Hammond, and C. Batten. Dotted-line extrapolations by C. Moore: Chuck Moore,
2011, “Data processing in exascale-class computer systems,” The Salishan Conference on High
Speed Computing, April 27, 2011.
42 © 2019 Arm Limited
Limits to Single Core Performance
• On-chip wiring
• Wire delays scale relatively poorly compared to logic delays.
• This limits the amount of state reachable in one clock cycle.
• Unfortunately, this limits the performance of large complex processors.

43 © 2019 Arm Limited

Specialization
What does specialization allow us to do?
• Remove infrequently used parts of the processor.
• Tune the instruction set for common operations or replace with hardwired control .
• Exploit forms of parallelism abundant in the application(s) – we often see a specialized
processing element and local memory reproduced many times.
• Instantiate specialized memories and tune their widths and sizes.
• Provide specialized interconnect between components.
• Optimize data-use patterns.

44 © 2019 Arm Limited

Specialization

Data assume a 45 nm process @0.9 V

M. Horowitz, Computing's energy problem (and what we can do about it), IEEE, March. 6, 2014.
[Online]. Available: https://ieeexplore.ieee.org/document/6757323 

45 © 2019 Arm Limited

Limits to Specialization
• There are costs associated with designing each new accelerator.
• The chip, or “ASIC,” produced may only be competitive in a smaller target market,
reducing profitability.
• Specialization reduces flexibility.
• The logic invested in specialized accelerators is no longer general-purpose.
• Algorithm changes may render specialized hardware obsolete.
• Once we’ve specialized, further gains may be difficult to achieve.
• Specialization isn’t immune to the concept of diminishing returns.

46 © 2019 Arm Limited