Table of Content  Branch Up  

Harvard vs. Von Newman

General-purpose machines followed the "stored program" approach since the early days. Code and data share the same addressable memory space, that is, code is taken into the CPU in the same manner as data is.

This is great for general-purpose machines, but the price to pay is the need for fetching the operational code for each instruction of the program, which takes precious bus cycles that obviously impact the overall performance.

For single-task computers (controllers) the program is fixed. We can take advantage of this by storing that program in a separate memory space. Now, the paths for code and data are separate so we don't need to fetch the operational code; we can just decode and execute instructions directly from where they are: each memory location (just to say) acts as the instruction register per se.

I'm thinking in a small program, of course, so by "memory" I mean a diodes-matrix (kind of old-fashioned ROM). This machine would consists of a few circuit boards, one of those being the diodes-matrix where the little control program resides.