From: Phil Karn <karn@qualcomm.com>
Date: Thu, 21 Apr 94 16:07:24 PDT
To: eb@sr.hp.com
Subject: Re: cryptophone ideas
In-Reply-To: <9404202202.AA18655@srlr14.sr.hp.com>
Message-ID: <199404212304.QAA21439@servo.qualcomm.com>
MIME-Version: 1.0
Content-Type: text/plain


>They are capable of doing 2 data moves, a 16x16 multiply, a 40 bit
>accumulate and a prefech of the next instruction all in 100ns.

This is where a DSP really shines, since it's the fundamental
operation in digital filtering; indeed it wouldn't be a DSP if it
couldn't do a multiply/accumulate in a single clock cycle.

But I wouldn't be too surprised if general purpose CPUs eventually get
the same capability. And once they are, the distinction between a
"DSP" and a "general purpose" CPU will pretty much vanish. DSPs are
notoriously harder to program than general purpose CPUs, and being
lower volume items they won't be able to compete in price or clock
speed with general purpose CPUs made in the millions.

But that's in the future. There's not much alternative to using a DSP
chip right now if you want high quality low bit rate speech, but
unfortunately the low-cost DSPs now appearing on PC sound cards are
not quite up to the task yet. I think CELP encoding requires something
like 30 million multiplies per second, which is beyond the reach of a
12.5 Mhz AD2105. On the other hand, simpler schemes and/or clever
coding tricks might make it possible. And since these boards are now
widely available in computer stores, they're hard to ignore in a
project like this. Has anybody looked at them in detail?

Phil