> > Me, too, but my impression from reading this list was that the problem was
> > the arm32 port's interrupt structure, and resulting interrupt response
> > latency [...]
>
> Herb's correct; it's the floating point performance.  We proved this with
> the June snapshot by replacing libc with the faster-fp version available
> on ftp.netbsd.org.

There were multiple problems.

The ones I talked about in earlier mail (a couple of months ago) were
with playing CD-rate audio directly, and had nothing to do with
floating point.  The problems I encountered had to do with interrupt
latency and possibly the speed of the DMA controller.  I haven't
tested it much since then, but I found at the time that limiting the
amount of network I/O made it work okay.

That said, the popular MPEG players do run slower than real-time due
to the FPU emulation.