Re: Structured Programming using Forth

On Apr 6, 4:59 pm, John Passaniti <n...@xxxxxxxxxxxxxxxxx> wrote:
Yes, and that's why I wrote that you seem to be confusing the width of
the A/D with the width that is needed. Yes, the dominate bit-width of
digital audio these days is 24 bits. But that is only the width used by
the A/D and D/A converters. Even the most basic digital audio system
will require more bits in the intermediate calculations used in home
theatre systems.

Absolutely. Errors accumulate in the lower bits of calculations and
so intermediate calculations may need a lot more than 24 bits and
this is only an 18-bit processor, although a 36-bit multiply step
will
be available in the instruction set. Since it takes multiple steps
and isn't a single cycle MAC multiple nodes will be needed to match
the speed of some DSP chips. But nodes are small and cheap and as
long as you end up with a significantly lower cost or power
consumption
for the calculation that needs to be done you have a win. So we
have compared the performance on these things to other solutions.

RF DSP needs to be considerably faster than audio processing but
has lower bit resolution requirements.

Everything I find on the Intellasys web site and elsewhere suggests that
the SEAforth machine is an 18-bit machine. The instruction set summary
that I see described doesn't have a multiply, but does have a multiply
step that is used to synthesize a multiply. So, when discussing how the
SEAforth chip will fare against DSPs, I can make the following
intelligent guesses:

1. Since the processor is likely an 18-bit machine, we're going to need
multi-precision math in order to do the various calculations used in
modern digital audio systems.

yes

2. Since the processor doesn't have a hardware multiply but does have a
multiply step (called "+*"), it's reasonable to assume that MAC
operations will be synthesized from more primitive operations.

Yes, there are many different representations of numbers used and
multiply operations are done by a sequence of steps. More steps are
required than with more complex chips, but the steps are faster.
Still one of these 20K cores may not be as fast at MAC as a DSP
that is ten times bigger, but ten cores might compare favorably.

We have done lots of real comparisons of this type to quantify
these things. Those are some of the bottom line things. You are
correct about the precision issues and performance issues.


.

Relevant Pages

  • Re: Structured Programming using Forth
    ... so intermediate calculations may need a lot more than 24 bits and ... Since it takes multiple steps ... I realized I could use multiple nodes to get higher performance and/or to do the kinds of SIMD instructions some DSPs offer. ... We're currently not very interested in RF applications other than for simple control applications. ...
    (comp.lang.forth)
  • Re: Caching: What do you do when it doesnt work?
    ... tm> (multiple 16-way units) unix system from an unnamed vendor. ... tm> calculations have failed? ...
    (comp.arch)
  • Re: New Mac Pros - 3Ghz
    ... I just opened Excel and it already ... Now that I've opened a worksheet, ... Tell me again that it can't make use of multiple processors... ... UI, so it doesn't lock and such, but actual calculations are all done ...
    (comp.sys.mac.advocacy)
  • Re: efficiency
    ... In one sense, it would be more efficient to do all the calculations in one query, so as to avoid writing multiple queries, and possibly the work entailed in stitching the correlated results together again later. ... "The Dummy in Access for Dummies" wrote: ...
    (microsoft.public.access.tablesdbdesign)