[Info-vax] OT: IA-128 ???

[glen herrmannsfeldt]

JF Mezei <jfmezei.spamnot at vaxination.ca> wrote:
> Marc Schlensog wrote:
 
>> A 128bit file system would be an entirely different story. But you
>> don't need a 128bit CPU for that.
 
> Lets say they were to produce an 8086 with 64 bit addressing, and 128
> bit registers and data paths.
 
> Are there many operations that were really benefit from that ?

No.  Note that 64 bit operations are available on 32 bit processors.
Add and subtract are easy.  Multiply a little harder, and divide
just a little bit harder than that, but all can be done.

32 bit floating point was easily done on the 8080.
 
> Moving long strings from one location to another would be faster since
> half the fetch/strore operations would be needed to move data around.

I believe that by now the processor can do it fast enough with
the current instructions that it is limited by the path between
cache and main memory.  That path could be made wider even on
a 32 or 64 bit processor.
 
> But for integer math, would it really make a difference ? How many
> applications need to process numbers greater than 2^64 ? I think that
> even Bill Gates's net worth fits comfortably within 64 bits.

Even for the ones that do, 32 bits is enough most of the time.
 
> Does IEEE already have a 128 version of the floating point
> representation ? Would creating 128but floating point registers provide
> advantages for high end computtional work (or video games) ?

In designing the 360/85 IBM did a study of the cost vs. value
of 128 bit floating point.  The result was that it was worthwhile
to do everything except divide in hardware.  Divide occurred rarely
enough that doing it in software was good enough.  Only very recently
was DXR (extended precision divide) added, late in the ESA/390 years.
 
> The move from 32 bit to 64 bit was made, as I recall, mostly to allow to
> go beyond the 4 gig of ram limit in 32 bit, and mostly to allow atabase
> engine to store more of s database in memory to increase performance.

Well, they had a 36 bit physical address bus starting with the
pentium-pro.  The data base people knew how to use that, though it
wasn't so easy to do.  I believe that in the end it wasn't so hard
to do so they just went to 64.  
 
> Perhaps going 128 bits addressing  might allow mapping
> of very big files to memory ? In other words, you may not have more
> than 64 bit's worth of phyical memory,  but your whole database of
> terabytes of records would be mapped to memory (think global section in
> VMS terns).
 
> Google might be a driver for this, when you consider it has huge
> databases (images for google erth street vie etc).

-- glen