The architectures of automated computing systems have evolved from the first mechanical calculators constructed nearly two centuries ago to the broad array of modern electronic computer technologies we use directly and indirectly every day. Along the way, there have been stretches of incremental technological improvement interspersed with disruptive advances that drastically altered the trajectory of the industry. We can expect these trends to continue in the coming years.

In the 1980s, during the early days of personal computing, students and technical professionals eager to learn about computer technology had a limited range of subject matter available for this purpose. If they had a computer of their own, it was probably an IBM PC or an Apple II. If they worked for an organization with a computing facility, they might have used an IBM mainframe or a Digital Equipment Corporation VAX minicomputer. These examples, and a limited number of similar systems, encompassed most people’s exposure to the computer systems of the time.

Today, numerous specialized computing architectures exist to address widely varying user needs. We carry miniature computers in our pockets and purses that can place phone calls, record video, and function as full participants on the internet. Personal computers remain popular in a format outwardly similar to the PCs of past decades. Today’s PCs, however, are orders of magnitude more capable than the early generations in terms of computing power, memory size, disk space, graphics performance, and communication ability. These capabilities enable modern PCs to easily perform tasks that would have been inconceivable on early PCs, such as the real-time generation of high-resolution 3D images.

Companies offering web services to hundreds of millions of users construct vast warehouses filled with thousands of tightly coordinated computer systems capable of responding to a constant stream of user requests with extraordinary speed and precision. Machine learning systems are trained through the analysis of enormous quantities of data to perform complex activities such as driving automobiles.

This chapter begins with a presentation of some key historical computing devices and the leaps in technology associated with them. We will then examine some significant modern-day trends related to technological advances and introduce the basic concepts of computer architecture, including a close look at the 6502 microprocessor and its instruction set. The following topics will be covered in this chapter:

• The evolution of automated computing devices
• Moore’s law
• Computer architecture

Files for this chapter, including answers to the exercises, are available at https://github.com/PacktPublishing/Modern-Computer-Architecture-and-Organization-Second-Edition.

This section reviews some classic machines from the history of automated computing devices and focuses on the major advances each embodied. Babbage’s Analytical Engine is included here because of the many leaps of genius represented in its design. The other systems are discussed because they embodied significant technological advances and performed substantial real-world work over their lifetimes.

Charles Babbage’s Analytical Engine

Although a working model of the Analytical Engine was never constructed, the detailed notes Charles Babbage developed from 1834 until his death in 1871 described a computing architecture that appeared to be both workable and complete. The Analytical Engine was intended to serve as a general-purpose programmable computing device. The design was entirely mechanical and was to be constructed largely of brass. The Analytical Engine was designed to be driven by a shaft powered by a steam engine.

Borrowing from the punched cards of the Jacquard loom, the rotating studded barrels used in music boxes, and the technology of his earlier Difference Engine (also never completed in his lifetime, and more of a specialized calculating device than a computer), the Analytical Engine’s design was, otherwise, Babbage’s original creation.

Unlike most modern computers, the Analytical Engine represented numbers in signed decimal form. The decision to use base-10 numbers rather than the base-2 logic of most modern computers was the result of a fundamental difference between mechanical technology and digital electronics. It is straightforward to construct mechanical wheels with 10 positions, so Babbage chose the human-compatible base-10 format because it was not significantly more technically challenging than using some other number base. Simple digital circuits, on the other hand, are not capable of maintaining 10 different states with the ease of a mechanical wheel.

All numbers in the Analytical Engine consisted of 40 decimal digits. The large number of digits was likely chosen to reduce problems with numerical overflow. The Analytical Engine did not support floating-point mathematics.

Each number was stored on a vertical axis containing 40 wheels, with each wheel capable of resting in 10 positions corresponding to the digits 0-9. A 41^st number wheel contained the sign: any even number on this wheel represented a positive sign, and any odd number represented a negative sign. The Analytical Engine axis was somewhat analogous to the register used in modern processors, except the readout of an axis was destructive—reading an axis would set it to 0. If it was necessary to retain an axis’s value after it had been read, another axis had to store a copy of the value during the readout. Numbers were transferred from one axis to another, or used in computations, by engaging a gear with each digit wheel and rotating the wheel to extract the numerical value. The set of axes serving as system memory was referred to as the store.

The addition of two numbers used a process somewhat similar to the method of addition taught to schoolchildren. Assume a number stored on one axis, let’s call it the addend, was to be added to a number on another axis that we will call the accumulator. The machine would connect each addend digit wheel to the corresponding accumulator digit wheel through a train of gears. It would then simultaneously rotate each addend digit downward to 0 while driving the accumulator digit an equivalent rotation in the increasing direction. If an accumulator digit wrapped around from 9 to 0, the next most significant accumulator digit would increment by 1. This carry operation would propagate across as many digits as needed (think of adding 1 to 999,999). By the end of the process, the addend axis would hold the value 0 and the accumulator axis would hold the sum of the two numbers. The propagation of carries from one digit to the next was the most mechanically complex part of the addition process.

Operations in the Analytical Engine were sequenced by music box-like rotating barrels in a construct called the mill, which is analogous to the control unit of a modern CPU.

Each Analytical Engine instruction was encoded in a vertical row of locations on the barrel, where the presence or absence of a stud at a particular location either engaged a section of the Engine’s machinery or left the state of that section unchanged. Based on Babbage’s estimate of the Engine’s execution speed, the addition of two 40-digit numbers, including the propagation of carries, would take about 3 seconds.

Babbage conceived several important concepts for the Engine that remain relevant to modern computer systems. His design supported a degree of parallel processing consisting of simultaneous multiplication and addition operations that accelerated the computation of series of values intended to be output as numerical tables. Mathematical operations such as addition supported a form of pipelining, in which sequential operations on different data values overlapped in time.

Babbage was well aware of the difficulties associated with complex mechanical devices, such as friction, gear backlash, and wear over time. To prevent errors caused by these effects, the Engine incorporated mechanisms called lockings that were applied during data transfers across axes. The lockings forced the number wheels into valid positions and prevented the accumulation of small errors from allowing a wheel to drift to an incorrect value. The use of lockings is analogous to the amplification of potentially weak input signals to produce stronger outputs by the digital logic gates in modern processors.

The Analytical Engine was to be programmed using punched cards and supported branching operations and nested loops. The most complex program intended for execution on the Analytical Engine was developed by Ada Lovelace to compute the Bernoulli numbers, an important sequence in number theory. The Analytical Engine code to perform this computation is recognized as the first published computer program of substantial complexity.

Babbage constructed a trial model of a portion of the Analytical Engine mill, which is currently on display at the Science Museum in London.

ENIAC

ENIAC, the Electronic Numerical Integrator and Computer, was completed in 1945 and was the first programmable general-purpose electronic computer. The system consumed 150 kilowatts of electricity, occupied 1,800 square feet of floor space, and weighed 27 tons.

The design was based on vacuum tubes, diodes, and relays. ENIAC contained over 17,000 vacuum tubes that functioned as switching elements.

Similar to the Analytical Engine, it used base-10 representation of 10-digit decimal numbers implemented using 10-position ring counters (the ring counter will be discussed in Chapter 2, Digital Logic).

Input data was received from an IBM punch-card reader and the output of computations was delivered by a card punch machine.

The ENIAC architecture was capable of complex sequences of processing steps including loops, branches, and subroutines. The system had 20 10-digit accumulators that functioned like registers in modern computers. It did not initially have any memory beyond the accumulators. If intermediate values were required for use in later computations, the data had to be written to punch cards and read back in when needed. ENIAC could perform about 385 multiplications per second.

ENIAC programs consisted of plugboard wiring and switch-based function tables. Programming the system was an arduous process that often took the team of talented female programmers weeks to complete. Reliability was a problem, as vacuum tubes failed regularly, requiring troubleshooting on a day-to-day basis to isolate and replace failed tubes.

In 1948, ENIAC was improved by adding the ability to program the system via punch cards rather than plugboards. This greatly enhanced the speed with which programs could be developed. As a consultant for this upgrade, John von Neumann proposed a processing architecture based on a single memory region holding program instructions and data, a processing component with an arithmetic logic unit and registers, and a control unit that contained an instruction register and a program counter. Many modern processors continue to implement this general structure, now known as the von Neumann architecture. We will discuss this architecture in detail in Chapter 3, Processor Elements.

Early applications of ENIAC included analyses related to the development of the hydrogen bomb and the computation of firing tables for long-range artillery.

IBM PC

In the years following the construction of ENIAC, several technological breakthroughs resulted in remarkable advances in computer architectures:

• The invention of the transistor in 1947 by John Bardeen, Walter Brattain, and William Shockley delivered a vast improvement over the vacuum tube technology prevalent at the time. Transistors were faster, smaller, consumed less power, and, once production processes had been sufficiently optimized, were much more reliable than the failure-prone tubes.
• The commercialization of integrated circuits in 1958, led by Jack Kilby of Texas Instruments, began the process of combining large numbers of formerly discrete components onto a single chip of silicon.
• In 1971, Intel began production of the first commercially available microprocessor, the Intel 4004. The 4004 was intended for use in electronic calculators and was specialized to operate on 4-bit binary-coded decimal digits.

From the humble beginnings of the Intel 4004, microprocessor technology advanced rapidly over the ensuing decade by packing increasing numbers of circuit elements onto each chip and expanding the capabilities of the microprocessors implemented on those chips.

The Intel 8088 microprocessor

IBM released the IBM PC in 1981. The original PC contained an Intel 8088 microprocessor running at a clock frequency of 4.77 MHz and featured 16 KB of Random Access Memory (RAM), expandable to 256 KB. It included one or, optionally, two floppy disk drives. A color monitor was also available. Later versions of the PC supported more memory, but because portions of the address space had been reserved for video memory and Read-Only Memory (ROM), the architecture could support a maximum of 640 KB of RAM.

The 8088 contained 14 16-bit registers. Four were general-purpose registers (AX, BX, CX, and DX). Four were memory segment registers (CS, DS, SS, and ES) that extended the address space to 20 bits. Segment addressing functioned by adding a 16-bit segment register value, shifted left by 4 bit positions, to a 16-bit offset contained in an instruction to produce a physical memory address within a 1 MB range.

The remaining 8088 registers were the Stack Pointer (SP), the Base Pointer (BP), the Source Index (SI), the Destination Index (DI), the Instruction Pointer (IP), and the Status Flags (FLAGS). Modern x86 processors employ an architecture remarkably similar to this register set (Chapter 10, Modern Processor Architectures and Instruction Sets, will cover the details of the x86 architecture). The most obvious differences between the 8088 and x86 are the extension of the register widths to 32 bits in x86 and the addition of a pair of segment registers (FS and GS) that are used today primarily as data pointers in multithreaded operating systems.

The 8088 had an external data bus width of 8 bits, which meant it took two bus cycles to read or write a 16-bit value. This was a performance downgrade compared to the earlier 8086 processor, which employed a 16-bit external bus. However, the use of the 8-bit bus made the PC more economical to produce and provided compatibility with lower-cost 8-bit peripheral devices. This cost-sensitive design approach helped reduce the purchase price of the PC to a level accessible to more potential customers.

Program memory and data memory shared the same address space and the 8088 accessed memory over a single bus. In other words, the 8088 implemented the von Neumann architecture. The 8088 instruction set included instructions for data movement, arithmetic, logical operations, string manipulation, control transfer (conditional and unconditional jumps, and subroutine call and return), input/output, and additional miscellaneous functions. The processor required about 15 clock cycles per instruction on average, resulting in an execution speed of 0.3 million instructions per second (MIPS).

The 8088 supported nine distinct modes for addressing memory. This variety of modes was needed to efficiently access a single item at a time as well as for iterating over sequences of data.

The segment registers in the 8088 architecture provided a seemingly clever way to expand the range of addressable memory without increasing the length of most instructions referencing memory locations. Each segment register allowed access to a 64-kilobyte block of memory beginning at a physical memory address defined at a multiple of 16 bytes. In other words, the 16-bit segment register represented a 20-bit base address with the lower four bits set to zero. Instructions could then reference any location within the 64-kilobyte segment using a 16-bit offset from the address defined by the segment register.

The CS register selected the code segment location in memory and was used in fetching instructions and performing jumps and subroutine calls and returns. The DS register defined the data segment location for use by instructions involving the transfer of data to and from memory. The SS register set the stack segment location, which was used for local memory allocation within subroutines and for storing subroutine return addresses.

Programs that required less than 64 kilobytes in each of the code, data, and stack segments could ignore the segment registers entirely because those registers could be set once at program startup (programming language compilers would do this automatically) and remain unchanged through execution. Easy!

Things got quite a bit more complicated when a program’s data size increased beyond 64 kilobyte. Though the use of segment registers resulted in a clean hardware design, using those registers caused many headaches for software developers. Compilers for the 8088 architecture distinguished between near and far references to memory. A near pointer represented a 16-bit offset from the current segment register base address. A far pointer contained 32 bits of addressing information: a 16-bit segment register value and a 16-bit offset. Far pointers consumed an additional 16 bits of data memory and they also required additional processing time.

Making single memory access using a far pointer involved the following steps:

1. Save the current segment register contents to a temporary location
2. Load the new segment value into the register
3. Access the data (reading or writing as needed) using an offset from the segment base
4. Restore the original segment register value

When using far pointers, it was possible to declare data objects (for example, an array of characters representing a document in a text editor) up to 64 KB in size. If you needed a larger structure, you had to work out how to break it into chunks no larger than 64 KB and manage them yourself. As a result of such segment register manipulations, programs that required extensive access to data items larger than 64 KB became quite complex and were susceptible to code size bloat and slower execution.

The IBM PC motherboard contained a socket for an optional Intel 8087 floating-point coprocessor. The designers of the 8087 invented data formats and processing rules for 32-bit and 64-bit floating-point numbers that became enshrined in 1985 as the IEEE 754 floating-point standard, which remains in near-universal use today. The 8087 could perform about 50,000 floating-point operations per second. We will look at floating-point processing in detail in Chapter 9, Specialized Processor Extensions.

The Intel 80286 and 80386 microprocessors

The second generation of the IBM PC, the PC AT, was released in 1984. AT stood for Advanced Technology, which referred to several significant enhancements over the original PC that mostly resulted from the use of the Intel 80286 processor.

Like the 8088, the 80286 was a 16-bit processor, and it maintained backward compatibility with the 8088: 8088 code could run unmodified on the 80286. The 80286 had a 16-bit data bus and 24 address lines supporting a 16-megabyte address space. The external data bus width was 16 bits, improving data access performance over the 8-bit bus of the 8088. The instruction execution rate (instructions per clock cycle) was about double the 8088 in many applications. This meant that at the same clock speed, the 80286 would be twice as fast as the 8088. The original PC AT clocked the processor at 6 MHz and a later version operated at 8 MHz. The 6 MHz variant of the 80286 achieved an instruction execution rate of about 0.9 MIPS, roughly three times that of the 8088.

The 80286 implemented a protected virtual address mode intended to support multiuser operating systems and multitasking.

In protected mode, the processor enforced memory protection to ensure one user’s programs could not interfere with the operating system or with other users’ programs. This groundbreaking technological advance in personal computing remained little used for many years, mainly because of the prohibitive cost of adding sufficient memory to a computer system to make it useful in a multiuser, multitasking context.

Following the 80286, the next generation of the x86 processor line was the 80386, introduced in 1985. The 80386 was a 32-bit processor with support for a flat 32-bit memory model in protected mode. The flat memory model allowed programmers to address up to 4 GB directly, without the need to manipulate segment registers. Compaq introduced an IBM PC-compatible personal computer based on the 80386 called the DeskPro in 1986. The DeskPro shipped with a version of Microsoft Windows targeted to the 80386 architecture.

The 80386 maintained substantial backward compatibility with the 80286 and 8088 processors. The processor architecture implemented in the 80386 remains the current standard x86 architecture. We will examine this architecture in detail in Chapter 10, Modern Processor Architectures and Instruction Sets.

The initial version of the 80386 was clocked at 33 MHz and achieved about 11.4 MIPS. Modern implementations of the x86 architecture run several hundred times faster than the original as the result of higher clock speeds, performance enhancements, including the extensive use of multilevel cache memory, and more efficient instruction execution at the hardware level. We will examine the benefits of cache memory in Chapter 8, Performance-Enhancing Techniques.

The iPhone

In 2007, Steve Jobs introduced the iPhone to a world that had no idea it had any use for such a device. The iPhone built upon previous revolutionary advances from Apple Computer, including the Macintosh computer, released in 1984, and the iPod music player of 2001. The iPhone combined the functions of the iPod, a mobile telephone, and an internet-connected computer.

The iPhone did away with the hardware keyboard that was common on smartphones of the time and replaced it with a touchscreen capable of displaying an onscreen keyboard, or any other type of user interface. In addition to touches for selecting keyboard characters and pressing buttons, the screen supported multi-finger gestures for actions such as zooming a photo.

The iPhone ran the OS X operating system, the same OS used on the flagship Macintosh computers of the time.

This decision immediately enabled the iPhone to support a vast range of applications already developed for Macs and empowered software developers to rapidly introduce new applications tailored to the iPhone, after Apple began allowing third-party application development.

The iPhone 1 had a 3.5” screen with a resolution of 320x480 pixels. It was 0.46 inches thick (thinner than other smartphones), had a built-in 2-megapixel camera, and weighed 4.8 oz. A proximity sensor detected when the phone was held to the user’s ear and turned off screen illumination and touchscreen sensing during calls. It had an ambient light sensor to automatically set the screen brightness and an accelerometer that detected whether the screen was being held in portrait or landscape orientation.

The iPhone 1 included 128 MB of RAM and 4 GB, 8 GB, or 16 GB of flash memory, and supported Global System for Mobile communications (GSM) cellular communication, Wi-Fi (802.11b/g), and Bluetooth.

In contrast to the abundance of openly available information about the IBM PC, Apple was notoriously reticent about releasing the architectural details of the iPhone’s construction. Apple released no information about the processor or other internal components of the first iPhone, simply referring to it as a closed system.

Despite the lack of official information from Apple, other parties enthusiastically tore down the various iPhone models and attempted to identify the phone’s components and how they interconnected. Software sleuths have devised various tests that attempt to determine the specific processor model and other digital devices implemented within the iPhone. These reverse engineering efforts are subject to error, so descriptions of the iPhone architecture in this section should be taken with a grain of salt.

The iPhone 1 processor was a 32-bit ARM11 manufactured by Samsung running at 412 MHz. The ARM11 was an improved variant of previous-generation ARM processors and included an 8-stage instruction pipeline and support for Single Instruction-Multiple Data (SIMD) processing to improve audio and video performance. The ARM processor architecture will be discussed further in Chapter 10, Modern Processor Architectures and Instruction Sets.

The iPhone 1 was powered by a 3.7 V lithium-ion polymer battery. The battery was not intended to be replaceable, and Apple estimated it would lose about 20 percent of its original capacity after 400 charge and discharge cycles. Apple quoted up to 250 hours of standby time and 8 hours of talk time on a single charge.

Six months after the iPhone was introduced, Time magazine named the iPhone the “Invention of the Year” for 2007. In 2017, Time ranked the 50 Most Influential Gadgets of All Time. The iPhone topped the list.

In the next section, we will examine the interplay of technological advances in computing over time and the underlying physical limits of silicon-based integrated circuits.

For those working in the rapidly advancing field of computer technology, it is a significant challenge to make plans for the future. This is true whether the goal is to plot your own career path or for a giant semiconductor corporation to identify optimal R&D investments. No one can ever be completely sure what the next leap in technology will be, what effects from it will ripple across the industry and its users, or when it will happen. One approach that has proven useful in this difficult environment is to develop a rule of thumb, or empirical law, based on experience.

Gordon Moore co-founded Fairchild Semiconductor in 1957 and was later the chairman and CEO of Intel. In 1965, Moore published an article in Electronics magazine in which he offered his prediction of the changes that would occur in the semiconductor industry over the next 10 years. In the article, he observed that the number of formerly discrete components, such as transistors, diodes, and capacitors, that could be integrated onto a single chip had been doubling approximately yearly and the trend was likely to continue over the next 10 years. This doubling formula came to be known as Moore’s law. This was not a scientific law in the sense of the law of gravity. Rather, it was based on an observation of historical trends, and he believed this formulation had some ability to predict the future.

Moore’s law turned out to be impressively accurate over those 10 years. In 1975, he revised the predicted growth rate for the following 10 years to double the number of components per integrated circuit every 2 years, rather than yearly. This pace continued for decades, up until about 2010. In more recent years, the growth rate has appeared to decline slightly. In 2015, Brian Krzanich, Intel CEO, stated that the company’s growth rate had slowed to doubling about every two and a half years.

Even though the time to double integrated circuit density is increasing, the current pace represents a phenomenal rate of growth that can be expected to continue into the future, just not quite as rapidly as it once progressed.

Moore’s law has proven to be a reliable tool for evaluating the performance of semiconductor companies over the decades.

Companies have used it to set goals for the performance of their products and to plan their investments. By comparing the integrated circuit density increases for a company’s products against prior performance, and against other companies, semiconductor executives and industry analysts can evaluate and score company performance. The results of these analyses have fed directly into decisions to invest in enormous new fabrication plants and to push the boundaries of ever-smaller integrated circuit feature sizes.

The decades since the introduction of the IBM PC have seen tremendous growth in the capabilities of single-chip microprocessors. Current processor generations are hundreds of times faster, operate natively on 32-bit and 64-bit data, have far more integrated memory resources, and unleash vastly more functionality, all packed into a single integrated circuit.

The increasing density of semiconductor features, as predicted by Moore’s law, has enabled these improvements. Smaller transistors run at higher clock speeds due to the shorter connection paths between circuit elements. Smaller transistors also, obviously, allow more functionality to be packed into a given amount of die area. Being smaller and closer to neighboring components allows the transistors to consume less power and generate less heat.

There was nothing magical about Moore’s law. It was an observation of the trends in progress at the time. One trend was the steadily increasing size of semiconductor dies. This was the result of improving production processes that reduced the density of defects, which allowed acceptable production yield with larger integrated circuit dies. Another trend was the ongoing reduction in the size of the smallest components that could be reliably produced in a circuit. The final trend was what Moore referred to as the “cleverness” of circuit designers in making increasingly efficient and effective use of the growing number of circuit elements placed on a chip.

Traditional semiconductor manufacturing processes have begun to approach physical limits that will eventually put the brakes on growth under Moore’s law. The smallest features on current commercially available integrated circuits are around 5 nanometers (nm). For comparison, a typical human hair is about 50,000 nm thick, and a water molecule (one of the smallest molecules) is 0.28 nm across. There is a point beyond which it is simply not possible for circuit elements to become smaller as the sizes approach atomic scale.

In addition to the challenge of building reliable circuit components from a small number of molecules, other physical effects with names such as Abbe diffraction limit become significant impediments to single-digit nanometer-scale circuit production.

We won’t get into the details of these phenomena; it’s sufficient to know the steady increase in integrated circuit component density that has proceeded for decades under Moore’s law is going to become a lot harder to continue over the coming years.

This does not mean we will be stuck with processors essentially the same as those that are now commercially available. Even as the rate of growth in transistor density slows, semiconductor manufacturers are pursuing several alternative methods to continue growing the power of computing devices. One approach is specialization, in which circuits are designed to perform a specific category of tasks extremely well rather than performing a wide variety of tasks merely adequately.

Graphics Processing Units (GPUs) are an excellent example of specialization. The original generation of GPUs focused exclusively on improving the speed at which three-dimensional graphics scenes could be rendered, mostly for use in video gaming. The calculations involved in generating a three-dimensional scene are well defined and must be applied to thousands of pixels to create a single frame. The process is repeated for each subsequent frame, and frames must be redrawn at a 60 Hz or higher rate to provide a satisfactory user experience. The computationally demanding and repetitive nature of this task is ideally suited for acceleration via hardware parallelism. Multiple computing units within a GPU simultaneously perform essentially the same calculations on different input data to produce separate outputs. Those outputs are combined to generate the entire scene. Modern GPU architectures have been enhanced to support other computing domains, such as training neural networks on massive amounts of data. GPU architectures will be covered in detail in Chapter 6, Specialized Computing Domains.

As Moore’s law shows signs of fading over the coming years, what advances might take its place to kick off the next round of innovations in computer architectures? We don’t know for sure today, but some tantalizing options are currently under intense study. Quantum computing is one example of these technologies. We will cover that technology in Chapter 17, Quantum Computing and Other Future Directions in Computer Architectures.

Quantum computing takes advantage of the properties of subatomic particles to perform computations in a manner that traditional computers cannot. A basic element of quantum computing is the qubit, or quantum bit. A qubit is similar to a regular binary bit, but in addition to representing the states 0 and 1, qubits can attain a state that is a superposition (or mixture) of the 0 and 1 states. When measured, the qubit output will always be 0 or 1, but the probability of producing either output is a function of the qubit’s quantum state prior to being read. Specialized algorithms are required to take advantage of the unique features of quantum computing.

Another future possibility is that the next great technological breakthrough in computing devices will be something that we either haven’t thought of or, if we have thought about it, we may have dismissed the idea out of hand as unrealistic. The iPhone, discussed in the preceding section, is an example of a category-defining product that revolutionized personal communication and enabled the use of the internet in new ways. The next major advance may be a new type of product, a surprising new technology, or some combination of product and technology. Right now, we don’t know what it will be or when it will happen, but we can say with confidence that such changes are coming.

The next section introduces some fundamental digital computing concepts that must be understood before we delve into digital circuitry and the details of modern computer architecture in the coming chapters.

The descriptions of a number of key architectures from the history of computing presented in the previous sections of this chapter included some terms that may or may not be familiar to you. This section will introduce the conceptual building blocks that are used to construct modern-day processors and related computer subsystems.

Representing numbers with voltage levels

One ubiquitous feature of modern computers is the use of voltage levels to indicate data values. In general, only two voltage levels are recognized: a low level and a high level. The low level is often assigned the value 0, and the high level is assigned the value 1.

The voltage at any point in a circuit (digital or otherwise) is analog in nature and can take on any voltage within its operating range. When changing from the low level to the high level, or vice versa, the voltage must pass through all voltages in between. In the context of digital circuitry, the transitions between low and high levels happen quickly and the circuitry is designed to not react to voltages between the high and low levels.

Binary and hexadecimal numbers

The circuitry within a processor does not work directly with numbers, in any sense. Processor circuit elements obey the laws of electricity and electronics and simply react to the inputs provided to them. The inputs that drive these actions result from the code developed by programmers and from the data provided as input to the program. The interpretation of the output of a program as, say, numbers in a spreadsheet, or characters in a word processing program, is a purely human interpretation that assigns meaning to the result of the electronic interactions within the processor. The decision to assign 0 to the low voltage and 1 to the high voltage is the first step in the interpretation process.

The smallest unit of information in a digital computer is a binary digit, called a bit, which represents a discrete data element containing the value 0 or 1. Multiple bits can be placed together to enable the representation of a greater range of values. A byte is composed of 8 bits placed together to form a single value. A byte is the smallest unit of information that can be read from or written to memory by most modern processors. Some computers, past and present, use a different number of bits for the smallest addressable data item, but the 8-bit byte is the most common size.

A single bit can take on two values: 0 and 1. Two bits placed together can take on four values: 00, 01, 10, and 11. Three bits can take on eight values: 000, 001, 010, 011, 100, 101, 110, and 111. In general, a group of n bits can take on 2ⁿ values. An 8-bit byte, therefore, can represent 2⁸, or 256, unique values.

The binary number format is not most people’s first choice when it comes to performing arithmetic. Working with numbers such as 11101010 can be confusing and error-prone, especially when dealing with 32- and 64-bit values. To make working with these numbers somewhat easier, hexadecimal numbers are often used instead. The term hexadecimal is often shortened to hex.

In the hexadecimal number system, binary numbers are separated into groups of 4 bits. With 4 bits in the group, the number of possible values is 2⁴, or 16. The first 10 of these 16 numbers are assigned the digits 0-9, and the last 6 are assigned the letters A-F. Table 1.1 shows the first 16 binary values starting at 0, along with the corresponding hexadecimal digit and the decimal equivalent to the binary and hex values:

Table 1.1: Binary, hexadecimal, and decimal numbers

The binary number 11101010 can be represented more compactly by breaking it into two 4-bit groups (1110 and 1010) and writing them as the hex digits EA. A 4-bit grouping is sometimes referred to as a nibble, meaning it is half a byte. Because binary digits can take on only two values, binary is a base-2 number system. Hex digits can take on 16 values, so hexadecimal is base-16. Decimal digits can have 10 values, and therefore decimal is base-10.

When working with these different number bases, it is easy for things to become confusing. Is a number written as 100 a binary, hexadecimal, or decimal value? Without additional information, you can’t tell. Various programming languages and textbooks have taken different approaches to remove this ambiguity. In most cases, decimal numbers are unadorned, so the number 100 is usually decimal. In programming languages such as C and C++, hexadecimal numbers are prefixed by 0x, so the number 0x100 is 100 hex. In assembly languages, either the prefix character $ or the suffix h might be used to indicate hexadecimal numbers. The use of binary values in programming is less common, mostly because hexadecimal is preferred due to its compactness. Some compilers support the use of 0b as a prefix for binary numbers.

Bits are numbered individually within a binary number, with bit 0 as the rightmost, least significant bit. Bit numbers increase in magnitude leftward, up to the most significant bit at the far left.

Some examples should make this clear. In Table 1.1, the binary value 0001b (1 decimal) has bit number 0 set to 1 and the remaining three bits are cleared to 0. For 0010b (2 decimal), bit 1 is set and the other bits are cleared. For 0100b (4 decimal), bit 2 is set and the other bits are cleared.

An 8-bit byte can take on values from $00h to $FF, equivalent to the decimal range 0-255. When performing addition at the byte level, the result can exceed 8 bits. For example, adding $01 to $FF results in the value $100. When using 8-bit registers, this represents a carry into the 9^th bit, which must be handled appropriately by the processor hardware and by the software performing the addition.

In unsigned arithmetic, subtracting $01 from $00 results in a value of $FF. This constitutes a wraparound to $FF. Depending on the computation being performed, this may or may not be the desired result. Once again, the processor hardware and the software must handle this situation to arrive at the desired result.

When appropriate, negative values can be represented using binary numbers. The most common signed number format in modern processors is two’s complement. In two’s complement, 8-bit signed numbers span the range from -128 to 127. The most significant bit of a two’s complement data value is the sign bit: a 0 in this bit represents a positive number and a 1 represents a negative number. A two’s complement number can be negated (multiplied by -1) by inverting all the bits, adding 1, and ignoring the carry. Inverting a bit means changing a 0 bit to 1 and a 1 bit to 0. See Table 1.2 for some step-by-step examples negating signed 8-bit numbers:

Table 1.2: Negation operation examples

Negating 0 returns a result of 0, as you would expect mathematically.

Table 1.3 shows how the binary values 00000000b to 11111111b correspond to signed values over the range -128 to 127, and unsigned values from 0 to 255:

Table 1.3: Signed and unsigned 8-bit numbers

Signed and unsigned representations of binary numbers extend to larger integer data types. 16-bit values can represent unsigned integers from 0 to 65,535, and signed integers in the range -32,768 to 32,767. 32-bit, 64-bit, and even larger integer data types are commonly available in modern processors and programming languages.

The 6502 microprocessor

This section introduces a processor architecture that is relatively simple compared to more powerful modern processors.

The intent here is to provide a whirlwind introduction to some basic concepts shared by processors spanning the spectrum from very low-end microcontrollers to sophisticated multi-core 64-bit processors.

The 6502 processor was introduced by MOS Technology in 1975. The 6502 found widespread use in video game consoles from Atari and Nintendo and in computers marketed by Commodore and Apple. Versions of the 6502 continue to be in widespread use today in embedded systems, with estimates of between 5 and 10 billion (yes, billion) units produced as of 2018. In popular culture, both Bender, the robot in Futurama, and the T-800 robot in The Terminator appear to have employed the 6502, based on onscreen evidence.

Like many early microprocessors, the 6502 was powered by 5 volts (V) direct current (DC). In these circuits, a low signal level is any voltage between 0 and 0.8 V. A high signal level is any voltage between 2 and 5 V. Voltages between these ranges occur only during transitions from low to high and from high to low. The low signal level is defined as logical 0, and the high signal level is defined as logical 1. Chapter 2, Digital Logic, will delve further into the electronic circuits used in digital electronics.

The word length of a processor defines the size of the fundamental data element the processor operates upon. The 6502 has a word length of 8 bits. This means the 6502 reads and writes memory 8 bits at a time and stores data internally in 8-bit wide registers.

Program memory and data memory share the same address space and the 6502 accesses its memory over a single bus. Like the Intel 8088, the 6502 implements the von Neumann architecture. The 6502 has a 16-bit address bus, enabling the addressing of 64 kilobytes of memory.

1 KB is defined as 2¹⁰, or 1,024 bytes. The number of unique binary combinations of the 16 address lines is 2¹⁶, which permits access to 65,536 byte-wide memory locations. Note that just because a device can address 64 KB, it does not mean there must be memory at each of those locations. The Commodore VIC-20, based on the 6502, contained just 5 KB of RAM and 20 KB of ROM.

The 6502 contains internal storage areas called registers. A register is a location in a logical device in which a word of information can be stored and acted upon during computation. A typical processor contains a small number of registers for temporarily storing data values and performing operations such as addition or address computations.

Figure 1.1 shows the 6502 register structure. The processor contains five 8-bit registers (A, X, Y, P, and S) and one 16-bit register (PC). The numbers above each register indicate the bit numbers at each end of the register:

Figure 1.1: 6502 register set

Each of the A, X, and Y registers can serve as a general-purpose storage location. Program instructions can load a value into one of those registers and, some instructions later, use the saved value for some purpose if the intervening instructions did not modify the register contents. The A register is the only register capable of performing arithmetic operations. The X and Y registers, but not the A register, can be used as index registers in calculating memory addresses.

The P register contains processor flags. Each bit in this register has a unique purpose, except for the bit labeled 1. The 1 bit is unused and can be ignored. Each of the remaining bits in this register is called a flag and indicates a specific condition that has occurred or represents a configuration setting. The 6502 flags are as follows:

• N: Negative sign flag: This flag is set when the result of an arithmetic operation sets bit 7 in the result. This flag is used in signed arithmetic.
• V: Overflow flag: This flag is set when a signed addition or subtraction results in overflow or underflow outside the range -128 to 127.
• B: Break flag: This flag indicates a Break (BRK) instruction has executed. This bit is not present in the P register itself. The B flag value is only relevant when examining the P register contents as stored on the stack by a BRK instruction or by an interrupt. The B flag is set to distinguish a software interrupt resulting from a BRK instruction from a hardware interrupt during interrupt processing.
• D: Decimal mode flag: If set, this flag indicates processor arithmetic will operate in Binary-Coded Decimal (BCD) mode. BCD mode is rarely used and won’t be discussed here, other than to note that this base-10 computation mode evokes the architectures of the Analytical Engine and ENIAC.
• I: Interrupt disable flag: If set, this flag indicates that interrupt inputs (other than the non-maskable interrupt) will not be processed.
• Z: Zero flag: This flag is set when an operation produces a result of 0.
• C: Carry flag: This flag is set when an arithmetic operation produces a carry.

The N, V, Z, and C flags are the most important flags in the context of general computing involving loops, counting, and arithmetic.

The S register is the stack pointer. In the 6502, the stack is the region of memory from addresses $100 to $1FF. This 256-byte range is used for the temporary storage of parameters within subroutines and holds the return address when a subroutine is called. At system startup, the S register is initialized to point to the top of this range. Values are “pushed” onto the stack using instructions such as PHA, which pushes the contents of the A register onto the stack.

When a value is pushed onto the stack, the 6502 stores the value at the address indicated by the S register, after adding the fixed $100 offset, and then decrements the S register. Additional values can be placed on the stack by executing more push instructions. As additional values are pushed, the stack grows downward in memory. Programs must take care not to exceed the fixed 256-byte size of the stack when pushing data onto it.

Data stored on the stack must be retrieved in the reverse of the order from which it was pushed onto the stack. The stack is a Last-In, First-Out (LIFO) data structure, meaning when you “pop” a value from the stack, it is the byte most recently pushed onto it. The PLA instruction increments the S register by 1 and then copies the value at the address indicated by the S register (plus the $100 offset) into the A register.

The PC register is the program counter. This register contains the memory address of the next instruction to be executed. Unlike the other registers, the PC is 16 bits long, allowing access to the entire 6502 address space.

Each instruction consists of a 1-byte operation code, called opcode for short, and may be followed by 0 to 2 operand bytes, depending on the type of instruction. After each instruction executes, the PC updates to point to the next instruction following the one that just completed. In addition to automatic updates during sequential instruction execution, the PC can be modified by jump instructions, branch instructions, and subroutine call and return instructions.

The 6502 instruction set

We will now examine the 6502 instruction set. Instructions are individual processor commands that, when strung together sequentially, execute the algorithm coded by the programmer. An instruction contains a binary number called an operation code (or opcode) that tells the processor what to do when that instruction executes.

If they wish, programmers can write code directly using processor instructions. We will see examples of this later in this section. Programmers can also write code in a so-called high-level language. The programmer then uses a software tool called a compiler that translates the high-level code into a (usually much longer) sequence of processor instructions.

In this section, we are working with code written as sequences of processor instructions. This form of source code is called assembly language.

Each of the 6502 instructions has a three-character mnemonic. In assembly language source files, each line of code contains an instruction mnemonic followed by any operands associated with the instruction. The combination of the mnemonic and the operands defines the addressing mode. The 6502 supports several addressing modes providing a great deal of flexibility in accessing data in registers and memory. For this introduction, we’ll only work with the immediate addressing mode, in which the operand itself contains a value rather than indicating a register or memory location containing the value. An immediate value is preceded by a # character.

In 6502 assembly, decimal numbers have no adornment (48 means 48 decimal), while hexadecimal values are preceded by a $ character ($30 means 30 hexadecimal, equivalent to 00110000b and to 48 decimal). An immediate decimal value looks like #48 and an immediate hexadecimal value looks like #$30.

Some assembly code examples will demonstrate the 6502 arithmetic capabilities. Five 6502 instructions are used in the following examples:

• LDA loads register A with a value
• ADC performs addition using Carry (the C flag in the P register) as an additional input and output
• SBC performs subtraction using the C flag as an additional input and output
• SEC sets the C flag directly
• CLC clears the C flag directly

Since the C flag is an input to the addition and subtraction instructions, it is important to ensure it has the correct value prior to executing the ADC or SBC instructions. Before initiating an addition operation, the C flag must be clear to indicate there is no carry from a prior addition. When performing multi-byte additions (say, with 16-bit, 32-bit, or 64-bit numbers), the carry, if any, will propagate from the sum of one byte pair to the next as you add the more significant bytes together. If the C flag is set when the ADC instruction executes, the effect is to add 1 to the result. After the ADC completes, the C flag serves as the ninth bit of the result: a C flag result of 0 means there was no carry, and a 1 indicates there was a carry from the 8-bit register.

Subtraction using the SBC instruction tends to be a bit more confusing to novice 6502 assembly language programmers. Schoolchildren learning subtraction use the technique of borrowing when subtracting a larger digit from a smaller digit. In the 6502, the C flag represents the opposite of Borrow. If C is 1, then Borrow is 0, and if C is 0, Borrow is 1. Performing a simple subtraction with no incoming Borrow requires setting the C flag before executing the SBC command.

The following examples employ the 6502 as a calculator using inputs defined as immediate values in the code and with the result stored in the A register. The Results columns show the final value of the A register and the states of the N, V, Z, and C flags:

Table 1.4: 6502 arithmetic instruction sequences

If you don’t happen to have a 6502-based computer with an assembler and debugger handy, there are several free 6502 emulators available online that you can run in your web browser. One excellent emulator is available at https://skilldrick.github.io/easy6502/. Visit the website and scroll down until you find a default code listing with buttons for assembling and running 6502 code. Replace the default code listing with a group of three instructions from Table 1.4 and then assemble the code.

To examine the effect of each instruction in the sequence, use the debugger controls to single-step through the instructions and observe the result of each instruction on the processor registers.

This section has provided a very brief introduction to the 6502 processor and a small subset of its capabilities. One point of this analysis was to illustrate the challenge of dealing with the issue of carries when performing addition and borrows when doing subtraction. From Charles Babbage to the designers of the 6502 to the developers of modern computer systems, computer architects have developed solutions to the problems of computation and implemented them using the best technology available to them.

This chapter began with a brief history of automated computing devices and described significant technological advances that drove leaps in computational capability. A discussion of Moore’s law followed with an assessment of its applicability over past decades and its implications for the future. The basic concepts of computer architecture were introduced through a discussion of the 6502 microprocessor registers and instruction set. The history of computer architecture is fascinating, and I encourage you to explore it further.

The next chapter will introduce digital logic, beginning with the properties of basic electrical circuits and proceeding through the design of digital subsystems used in modern processors. You will learn about logic gates, flip-flops, and digital circuits including multiplexers, shift registers, and adders. The chapter includes an introduction to hardware description languages, which are specialized computer languages used in the design of complex digital devices such as computer processors.

1. Using your favorite programming language, develop a simulation of a single-digit decimal adder that operates in the same manner as in Babbage’s Analytical Engine. First, prompt the user for two digits in the range 0-9: the addend and the accumulator. Display the addend, the accumulator, and the carry, which is initially 0. Perform a series of cycles as follows:
   1. If the addend is 0, display the values of the addend, accumulator, and carry and terminate the program
   2. Decrement the addend by 1 and increment the accumulator by 1
   3. If the accumulator is incremented from 9 to 0, increment the carry
   4. Go back to step 1
   5. Test your code with these sums: 0+0, 0+1, 1+0, 1+2, 5+5, 9+1, and 9+9
2. Create arrays of 40 decimal digits each for the addend, accumulator, and carry. Prompt the user for two decimal integers of up to 40 digits each. Perform the addition digit by digit using the cycles described in Exercise 1 and collect the carry output from each digit position in the carry array. After the cycles are complete, insert carries, and, where necessary, ripple them across digits to complete the addition operation. Display the results after each cycle and at the end. Test with the same sums as in Exercise 1 and also test the sums 99+1, 999999+1, 49+50, and 50+50.
3. Modify the program of Exercise 2 to implement the subtraction of 40-digit decimal values. Perform borrowing as required. Test with 0-0, 1-0, 1000000-1, and 0-1. What is the result for 0-1?
4. 6502 assembly language references data in memory locations using an operand value containing the address (without the # character, which indicates an immediate value). For example, the LDA $00 instruction loads the byte at memory address $00 into A. STA $01 stores the byte in A into address $01. Addresses can be any value in the range of 0 to $FFFF, assuming memory exists at the address and the address is not already in use for some other purpose. Using your preferred 6502 emulator, write 6502 assembly code to store a 16-bit value in addresses $00-$01, store a second value in addresses $02-$03, then add the two values and store the result in $04-$05. Be sure to propagate any carry between the two bytes. Ignore any carry from the 16-bit result. Test with $0000+$0001, $00FF+$0001, and $1234+$5678.
5. Write 6502 assembly code to subtract two 16-bit values in a manner similar to Exercise 4. Test with $0001-$0000, $0001-$0001, $0100-$00FF, and $0000-$0001. What is the result for $0000-$0001?
6. Write 6502 assembly code to store two 32-bit integers to addresses $00-03 and $04-$07, and then add them, storing the results in $08-$0B. Use a looping construct, including a label and a branch instruction, to iterate over the bytes of the two values to be added. Search the internet for the details of the 6502 decrement and branch instructions and the use of labels in assembly language. Hint: The 6502 zero-page indexed addressing mode works well in this application.

Join our community Discord space

Join the book’s Discord workspace for a monthly Ask me Anything session with the author: https://discord.gg/7h8aNRhRuY