Re: Superstitious learning in Computer Architecture

Steve Richfie1d wrote:
I have posted a
paper giving my own (warped) version of computer history that explains
the present architectural mess!

It certainly does make sense that, for example, once one company trying
to build a wafer-scale device failed, then investors would be hesitant
to back any project that involved wafer-scale technology... and that
investors would be less technically-qualified to asess the differences
between one proposal and another than the engineers - who, not having
the money, do not make the decisions.

My own pet peeve is the popularity of RISC. A simple, clean, CISC
design would save having to have two instructions for many common
operations - but, instead, our only two choices are RISC or the
baroque, legacy-rich, x86 architecture. (Needless to say, I mourn the
loss of the 68K architecture.)

But while I find the x86 to be baroque, because of its growth from a
16-bit architecture designed to "look like" an 8-bit architecture, I
felt some of your criticisms were a bit harsh. Instead of comparing it
to the STRETCH, I would compare it to the more successful 360/195 in
terms of computational design.

It is true that representing numbers by their logarithms makes for
efficient multiplication and division - and that addition and
subtraction are also possible, by using a lookup table for (r+1)/r (to
add a and b, where a<b, calculate r=a/b). I remember reading, at the
instigation of my professor, the original paper on the FOCUS system
using this method.

The problem is, though, that the precision to which one can do
arithmetic in such a fashion is bounded... and for many problems, the
24 or so bits of precision in a single-precision floating-point number
is not enough.

Of course, if one can add and multiply exactly, one can do
multi-precision arithmetic, and approaches such as SRT division involve
working on more than one bit at a time. Perhaps some technique based on
logarithmic representation could allow us to divide longer numbers up
into, say, 12-bit chunks that could be processed rapidly as individual
"digits", where one might allocate extra bits to the coded version of
the number to ensure that every possible 12-bit value had an _exact_
representation.

John Savard

.

Relevant Pages

  • Re: Why Cant Public Wi-Fi Be Made Secure?
    ... surface in Agha's architecture might desert some entire tunes. ... it'll most top the representation. ... She wants to rebuild red revenges as opposed to Cristof's pen. ...
    (sci.crypt)
  • Re: Is UINT_MAX required to be 2^n-1?
    ... Isn't there a requirement that the representation for a signed ... > integer type is a subrange of the corresponding unsigned integer type, ... As far as that goes, a BCD architecture is perfectly allowed, just ...
    (comp.std.c)
  • Re: Macro for supplying memset with an unsigned char
    ... non twos-complement architectures can only be found in museums ... representation as the target architecture. ... the preprocessor int types be the same as the actual C int types ...
    (comp.lang.c)
  • Write Great Code is now Available
    ... Chapter 4: Floating Point Representation ... Chapter 6: Memory Organization and Access ... Composite Data Types and Memory Objects ... Chapter 10: Instruction Set Architecture ...
    (alt.lang.asm)