18b computers (was: something much longer and unrelated)
frustum at pacbell.net
Fri May 22 01:51:34 CDT 2009
[changed the subject line due to drift]
bfranchuk at jetnet.ab.ca wrote:
> Grumbles at the lack of 9 bit wide ram chips
If solving that problem is a hurdle, you have much bigger problems that
you don't know about yet. :-)
>... No matter how you look at it
> any kind of 18 bit computer has power compared to a 16 bit or less cpu.
I think you are in love with the idea more than an actual computer.
It is trivial to imagine an anemic 18b computer that doesn't compare
well to a decent 16 bitter.
I worked on/helped design a CPU that used 9b bytes, with a 72b word
size. We sold a few million chips, which in the market we were in was a
miserable failure (this was in the mid 90s). The company was Chromatic
Research, and the first chip was called MPACT 1. That was followed up a
couple years later with MPACT 2.
The 9 bittedness came about because our memory system was based on
Rambus memory. We were one of the early adopters, due to the fact that
Chromatic Research and Rambus had a common founder, Mike Farmwald. At
the time, all Rambus parts were x9, so the 9th bit was supposedly "free"
and we might was well take advantage of it.
Most operations had an option of operating on eight 9b fields, four 18b
fields, or two 36b fields. The video system had a 666 pixel format.
MPEG en/decode was viewed as an important (primary even) feature, and
forming differences of two 8b components leads to a signed 9b number, so
it seemed like a good idea at the time.
The 9b aspect was a real burden that made many different aspects of the
design more complicated, sometimes in small ways, sometimes in large
ways. The main nice thing about it was it did give a bit of luxury to
encoding the instruction set.
One of the bigger pains caused by the 9b factor was that Rambus memory
carried a huge premium over regular DRAM, but it wasn't possible to
design out the Rambus interface.
Another problem caused by the x9 organization was that SIMD partitioning
of the shifter and generating the corresponding control terms is vastly
simpler for a power of two width than for one that is a multiple of 9.
In the end, the 9b aspect was a minor irritation, and the company
crashed down for a host of other much more significant reasons. Still,
my point is that having been there, 18b isn't somehow magically better
than 16b, and comes with its own drawbacks.
More information about the cctalk