Subject: Re: [linux-audio-dev] float a/d (was Linux support for IEEE1394 mLAN?)
From: Jussi Laako (jussi_AT_jlaako.pp.fi)
Date: Wed Oct 04 2000 - 20:59:43 EEST
Paul Barton-Davis wrote:
>
> >signal processing is memory bandwidth. It's the major limiting factor.
> for off-line processing, this might be an issue. however, for
> real-time applications, its not the critical limitation. and from my
> perspective, real-time work is by far the more interesting area.
I think just the opposite. When you are processing off-line you have almost
unlimited amout of time available, which is not the case for realtime
processing.
When using current (esp. Intel) CPUs for DSP, the main limiting factor is
memory bandwidth. You get almost no performance gain from increased clock
speed. And when using Intel CPU's for small size FFTs the speed is orders of
magnitude higher when using float vs double. Same thing with FIR filters.
Athlon has much smaller difference (around 1.5). Intel's PII and PIII
processor's L2 cache doesn't help FP processing "at all", only L1 (when
something is small enough to fit there). Athlon's L2 cache works well with
FP.
- Jussi Laako
-- PGP key fingerprint: 161D 6FED 6A92 39E2 EB5B 39DD A4DE 63EB C216 1E4B Available at: ldap://certserver.pgp.com, http://keys.pgp.com:11371
This archive was generated by hypermail 2b28 : Wed Oct 04 2000 - 21:44:23 EEST