Did he write it with fp (fp-ide) turbo pascal originally?
That is a completely ridiculous notion. Sorry, but Don Knuth does not use turbo pascal. Don uses real computers. Turbo Pascal runs on crappy toys.
The first version of TeX was written in SAIL, the famous language used by the DEC-10 (a mainframe made by Digital Equipment in the 60s and 70s). The language was designed for implementing artificial intelligence applications at Stanford (thence the name). The next version was written in "literate programming": The source code is written in a language called Web (it has nothing to do with the world-wide web, and predated it by about 20 years). You can take the Web source code and run it through either of two programs, one called "tangle", the other called "weave":
- The output of one of them is TeX code, which you can run through TeX, and it produces the full documentation for the source code. The documentation is so beautiful and clear, it reads like a book. Matter-of-fact, the source code was published in book form (I think it is roughly half a dozen volumes for TeX and MetaFont).
- The output of the other program is Pascal source code (in a strange dialect of Pascal which is specific to Knuth), and which you can compile into the TeX executable. Since the original Pascal compiler for the DEC-10 doesn't exist any longer, it is today typically converted into C code.
The other thing one has to remember that TeX is really not a document description language. It is actually a full programming language. Don Knuth is not a programmer, he is a computer scientist (and one of the greatest ones). You can actually implement a BASIC interpreter in TeX if you want (it's been done). The thing about TeX is: while it is a full programming language, it happens to be designed to produce DVI output as a side effect, so it happens to be most useful (and most easy to use) to write documents.
TeX itself is amazingly stable; it has had very few changes in the last dozen years or so. Its version number is asymptotically approaching pi.
In practice, raw TeX is an somewhat inefficient way to write documents. That's why people use LaTeX instead. LaTeX is implemented in TeX (it is nothing but a macro package for TeX). The huge difference is that TeX is a typesetting language: intellectually it thinks about "put this glyph here on the page", or "lay out the characters on this line while keeping sentences together but have nice line breacks". In contrast, LaTeX is a markup language (like SGML and HTML): intellectually it thinks about "this is a heading, and this is the caption under a figure". For large documents, LaTeX is much easier to use. Because LaTeX is nothing but a macro package for TeX, you have the whole power of TeX available, in addition to the document structuring mechanisms. LaTeX is the product of another computer science genius, Leslie Lamport (of Paxos fame). Obviously, Don and Leslie are friends, and live close to each other (Don is a professor at Stanford in Palo Alto, and Leslie worked for a long time a Digital Research in Palo Alto, and I think he now works at Microsoft in Mountain View or Sunnyvale).
To a large extent, TeX (and LaTeX) have nothing to do with postscript and PDF. They generate DVI output, which is device independent (thence the name). That output is then merged with fonts (which are often created by Knuth's program MetaFont, although people often use inferior Postscript fonts), to create device-specific output. The DVI format is relatively simple, and a DVI driver can be pretty short. I happen to know that the DVI driver for an early laser printer was written in 6 or 7 lines of APL ... but the lines were about 300 characters long (and APL source code is amazingly dense and bizarre, de-facto unreadable). I started using TeX in about 1982 (on an IBM mainframe, where for previewing we used Tektronix 4014 graphics terminals, and we had a wet toner-based electrostatic Versatec plotter for output). Today, most output from TeX goes into postscript or PDF format, but the original dvips driver (written by my former colleague Tom Rokicki from HP Labs) is small and efficient, and I'm sure you can find the source code for it on the web.
I'm sure a minimal TeX / LaTeX / MetaFond / dvips distribution could be built from original sources, and stored in a very small amount of space (dozens or hundreds of megabytes, mostly for the fonts). It would be an interesting exercise in retrocomputing. Would probably take a real expert a week or two to do; configuring and setting it up is famously difficult. Generating the fonts with mf takes considerable CPU time (we used to run the font generation overnight, when computers were still measured in single-digit MIPS).
Little personal story: One time I was at a meeting, sitting next to Leslie Lamport. He was crying. Don Knuth was also there, and he was also crying. So was I. Matter-of-fact, everyone in the room was crying. It was the memorial service for a colleague of ours, and the widow of our colleague was giving the eulogy on stage (she is a well-known computer science prof at Harvard herself). Very sad.