27 Feb
2009
27 Feb
'09
8:41 a.m.
On 27 Feb 2009 at 10:45, Paul Koning wrote:
Chuck> Maybe a 72 bit word length might have made
more sense, as that
Chuck> would have matched the 7090 double-precision word size. On
Chuck> the other hand, everything about the 6000 architecture does
Chuck> "fit together" nicely.
Right. For example, 15 and 30 bit instructions pack well, while a
72-bit word would not have produced that nice packing (and 15 bits is
the right basic instruction size). Also, 60 bits is the single
precision float size, while double is 120 bits -- so you get a lot
more precision than the 7090.
(DP on 7090 = 54 mantissa bits; SP on the 6000 = 48 mantissa bits. I
haven't compared the innards of the FP implementations on both
machines in detail; it might be that the 6000 preserved more
significance during calculcations than the 7090 DP).
It's always been my understanding that the 6600 took square aim at
IBM's hugely popular "scientific" machines, the 7090/7094. The
floating-point instructions have enough similarity to the 7090's to
support that contention, though the architecure in general doesn't.
A 72-bit word would have 4 18-bit parcels (allowing for 16 A, B and X
registers) or 2 36-bit parcels (allowing for a larger address space),
so it's not outside reason. Perhaps adding 12 bits to the word size
may have been too difficult or expensive from an engineering
standpoint, or simply not been cost-effective.
--Chuck