Computer Architecture

Computer Architecture

A set of rules and methods that describe the functionality, organisation, and implementation of computer systems, which can be applied at many levels/or layers such as:

• Processor architecture, memory architecture, instruction set architecture (ISA), etc.

What is Computer Architecture at the System Level?

how we link processors to devices for input and output, computer networks and other systems (e.g. architecture of the Internet), etc

What is current Computer Architecture standardised on?

von neumann architecture

Von Neumann Architecture

• Memory that stores data and instructions together

• A control unit that contains an instruction register and program counter

• A processing unit that contains an arithmetic logic unit (ALU) and processor registers 

• Input and output mechanisms

• The central processing unit (CPU) consisting of the ALU + CU often also contain high-speed cache memory

Harvard Architecture

• Instructions memories and data memories are separate, to overcome the bottleneck of Von Neumann Architecture

• Parallel (same time) access to instruction and data memory, can be faster

• Better cyber resilience against potential cyber attacks - more memory stores - if one gets attacked then not all data is lost - there is less probability of both stores getting attacked successfully

Modified Harvard Architecture

• Separates instruction and data caches internally

• But a single unified main memory is still visible to users/ programs

• Used in chips such as ARM9, MIPS, PowerPC, x86

• Form a users/programmers view, looks like Von Neumann Architecture

Speed Metrics

• Clock rate

  • E.g. a 1.87GHz processor makes 1.87 billion ticks per second

  • But, different instructions may take different numbers of ticks – sometimes unfair as a comparing metric

• Millions of instructions per second (MIPS)

  • a better indication of speed, but it depends on which instructions are counted (number of instructions)

  • Different results for different programs – again could be unfair

• Floating point operations per second (FLOPS)

  • Arguably a better indication of speed “where it counts” – again maybe unfair 

None are ideal. Also, don’t take into account input/output speed

Limiting Factors on Speed

Density limitations

• Number of transistors per square inch

• “Moore’s Law” (1965, updated 1975): transistor number on a silicon doubles every 2 years 

Power limitations (critical challenge)

• Around 1/3 of the power used to propagate the clock signal around the processor

• So, power and heat problems increase as clock rate increases – Cooling becomes very challenging to accommodate this problem

Coarser-grained Parallelism: Clustering

We can increase performance by linking computers using high-speed networks:

Leads to idea of “blade servers”

• Obviously they don’t all need screens, etc

Applications run across the cluster (ideally)

• Although, some applications can’t easily be decomposed in this way