Word

In the widest sense, a word is an ordered group of bits considered together. In computer architecture, a word may denote data processed by the CPU, an instruction, an address, or the unit by/in which main memory is accessed. In early and simple computers, these were to a large extent the same width. However, contemporary practice is generally to give each byte in main memory its own address. Instructions are also now often smaller than the ALU width; physical and virtual addresses are different sizes, etc.

Words are now usually a power-of-two multiple of an 8-bit byte in size, but in early computers, other word sizes were common. 36 bits was a very popular size (for reasons of accuracy of physical computations, the principal early application for computers) on early mainframes, such as the IBM 704 and its descendants; the Honeywell 6000 series‎‎; and the PDP-10 family. However, small word sizes were common too, e.g. the 12-bit words of the PDP-8 family.

Computer 'widths'

Machines' 'widths' are often considered to be the machine's word-size, but the discussion is complicated, and made even more confusing, because there is no universally accepted standard for what sets a machine's 'width'. Indeed, for most suggested benchmarks, one can find a machine which violates it:

• address space (KA10; 18 bits - 36-bit machine)
• instruction size (Motorola M68000 family; 16 bits - 32-bit machine)
• bus width (Motorola MC68000; 16 bits - 32-bit machine)
• ALU width (LSI-11; 8 bits - 16-bit machine)

Thus, the 'word' size is now usually used as the measure of the machine's 'size'.