COMPUTER ARCHITECTURE

1
COMPUTER ARCHITECTURE

• By Maria Ramila I. Jimenez
• XUCC

2
A CLASSIFICATION OF COMPUTER ARCHITECTURES

• Outline
• 1.2.1 Von Neumann Machines
• 1.2.2 Non-von Neumann Machines
• 1.3 Measuring the Quality of a Computer
  Architecture
• 1.4 Factors Influencing the Success of a
  Computer Architecture

3
Flynns classification of the 3 types of
computers in existence

• Flynn(1966) classified computer architectures by
  a variety of characteristics
• number of processors
• number of programs they can execute
• memory structure

4
Flynns classification of the 3 types of
computers in existence

• Categories
• Single instruction stream, single data stream
  (SISD)

5
Flynns classification of the 3 types of
computers in existence

• Single instruction stream, single data stream
  (SISD)
• SISD computers have one CPU that executes one
  instruction at a time (hence a single instruction
  stream) and fetches or store one item of data at
  a time( hence a single data stream).

6
Flynns classification of the 3 types of
computers in existence

• Single instruction stream, multiple data stream
  (SIMD)
• SIMD machines have a CU that operates like a von
  Neumann machine(i.e., it executes a single
  instruction stream), but SIMD machine have more
  than one PE. The CU generates the control signals
  for all of the PEs, which execute the same
  operation, generally in lockstep, on different
  data items (hence multiple data streams).

7
Flynns classification of the 3 types of
computers in existence

• Categories
• Single instruction stream, multiple data stream
  (SIMD)

8
Flynns classification of the 3 types of
computers in existence

• Multiple instruction stream, multiple data stream
  (MIMD)
• These machines have more than one independent
  processor, and each processor can execute a
  different program (hence multiple instruction
  streams) on its own data (hence multiple data
  streams).

9
Flynns classification of the 3 types of
computers in existence

• Multiple instruction stream, multiple data stream
  (MIMD)

10
(MISD)

• Multiple instruction stream, single data stream
  (MISD)
• Logically, machines in this class would execute
  several different programs on the same data item.
  There are currently no machines of this type,
  although some MIMD systems may be used in this
  manner.

11
Multiprocessor Architectures

• Multiprocessor architectures can be divided into
  2 categories based on the organization of their
  memory system
• Global memory (GM) system architectures.
• One global memory system is shared by all of the
  processors.
• Local-memory (LM) system architectures.
• One storage system exists for each processor.
  Multiprocessors with LM may also have GM and are
  also called multiple processors.

12
SIMD and MIMD

• SIMD and MIMD machines are parallel processors
  because they operate in parallel on more than one
  datum at a time.
• SIMD computers have the ff. characteristics
• They distribute processing over a large amount of
  hardware.
• They operate concurrently on many diff. data
  elements.
• They perform the same computation on all data
  elements.

13
SIMD and MIMD

• SIMD machines differ in the way their PEs access
  memory. The PEs of GM-SIMD computers share the
  same storage system, while those of LM-SIMD
  computers have independent storage systems.
• Processor arrays are SIMD architectures. They
  have 1 CU and many PEs. The CU generates control
  signals for all the PEs, w/c perform exactly the
  same computation simultaneously but w/ diff.
  data.

14
An LM-SIMD processor array
Processor Array
15
SIMD and MIMD

• The term CU in the SIMD context
• Usually the CU is itself a von Neumann computer
  with its own register set, ALU, and CU of the
  type to be described in chapter2. This computer
  is referred to as a CU bec. it was designed
  solely to control the PEs in a processor array,
  rather than to operate as a stand-alone computer.
  

16
SIMD and MIMD

• Historically, SIMD architectures include several
  components in addition to the CU and PEs. They
  usually include a host computer for
• loading programs,
• for configuring the array of PEs,
• and for supervising I/O.
• The host is usually a conventional computer w/ a
  well-established OS. The architects may also
  provide specialized I/O devices for

17
SIMD and MIMD

• reformatting the arrays of data, for example, or
  for performing high-speed I/O.
• MIMD computers have the ff. characteristics
• They distribute processing over a number of
  independent processors.
• They share resources, including main memory,
  among component processors.
• Each processor operates independently and
  concurrently.
• Each processor runs its own program.

18
SIMD and MIMD

• Different MIMD architectures have diff.
• interconnection networks,
• diff. processors,
• diff. memory-addressing structures, and
• diff. synchronization and control structures.
• Multiple-processor computers can be categorized
  as being either
• - tightly coupled
• - loosely coupled

19
SIMD and MIMD

• depending on how the processors access each
  others memory.
• The processors in a tightly coupled
  multiprocessor generally share one memory system.
• Those of a loosely coupled multiprocessor may
  also share a memory system, but each processor
  also has its own local memory and generally
  executes programs out of it.

20
Tightly coupled multiple-processor

• Architecture of a tightly coupled multiple
  processor computer

21
Loosely coupled multiple-processor computer

• Architecture of a loosely coupled
  multiple-processor computer.

22
SIMD and MIMD

• Thus tightly coupled and loosely coupled
  computers correspond approximately to the GM-MIMD
  and LM-MIMD classifications, respectively.
• Examples of GM-MIMD
• dual processor of the CDC 6600 series and Cray
  XM-P
• Examples of the LM-MIMD
• Carnegie-Mellon Cm and the Tandem/16.

23
Architectural categories
24
Other architectures

• In von Neumann machines programs determine the
  flow of control.
• In dataflow architectures, the availability of
  data determines when the machines will perform
  operations
• The computational models that dataflow processors
  implement, called dataflow models, are inherently
  parallel, and architects have designed dataflow
  machines to implement these models efficiently.

25
Other architectures

• A dataflow machine, like an MIMD computer has
  many independent PEs, but these do not execute
  programs. When the supervising processor finds
  available data, it routes the data and an
  instruction to a PE, w/c performs the operation
  and forwards the result to memory. Bec. There is
  no program in the conventional sense, dataflow
  architectures are neither SIMD nor MIMD.

26
Other architectures

• Neural networks are loosely based on biological
  systems. Like dataflow processors, they do not
  execute conventional programs and therefore do
  not fall into the SIMD or MIMD classifications.
  Current applications include
• signal processing
• pattern recognition

27
Other architectures

• Special-purpose machines
• Because of the specialized functions they
  perform. Generally, they use conventional
  architectures that have been optimized for
  specific applications. They differ from
  conventional architectures because they are
  required to solve a particular equation or to
  handle specific applications, such as
• very large arrays
• very large databases
• highly parallel algorithms

28
Other architectures

• Included in this group are
• artificial intelligence machines
• high-level-language machines
• image-processing machines
• three-dimensional display processors
• computers for embedded control

29
Other architectures

• Image-processing involves extracting info. from
  images or enhancing the quality of an image.
• three-dimensional display processors generally
  provide special-purpose hardware for displaying
  graphical images and find applications in CAD,
  CAM, and animation
• computers for embedded control