[Asterisk-Users] Hardware Transcoder

[John Todd]

At 6:05 PM -0700 on 6/11/04, George Pajari wrote:
>  > I have done a quick search and there are some nice looking
>>  dsp-pci cards out there. (Dunno abt prices). It may take
>>  some coding to get them working with Asterisk , and one
>>  would not require a super-power quad xeon processor if it
>>  had a huge dsp card.
>
>Perhaps but the fundamental premise upon which the Asterisk/Digium hardware
>is predicated is that commodity MIPS (i.e. generic server systems) are much
>cheaper than speciality DSP MIPS -- which is why the architecture used by
>Asterisk/Digium is to use board that are trivially simple (i.e. merely
>interface between the analog/digital trunks and the PCI bus) leaving all the
>signal processing to the server CPU. DSPs may be more "efficient" in some
>sense but because of the incredible volume of generic servers being shipped
>and the competitive pressure on pricing, they turn out to have better
>price/performance. DSP boards ship in volumes orders of magnitude smaller
>which makes them much more expensive.
>
>If power/density is an issue and you don't need the PCI slots to interface
>to digital trunks, take a look at using blade servers.
>
>g.

I make the argument quite often that using generic processors for 
DSP-like functions is obviously the way of the future for generic 
tasks.  As an example, processing power in MIPs vs. price for Intel 
or Intel-like processors looks pretty good when viewed from a 
distance - those curves seem to favor moving pretty much everything 
but the most exotic DSP work into the central CPU.  I firmly believe 
that telephony is a software problem now, not a hardware problem. 
"The DSP has died - long live dsp.c!"

However...

For transcoding and echo cancellation, we're not there yet for the 
high end telephony applications.  For Asterisk, this is OK, since 
most people on this list don't run "high end applications" with their 
Asterisk servers.  They have 1, or 5, or even 100 extensions on a PBX 
replacement, which works great - I don't consider that "high end", 
and therefore Asterisk works well for it's primary tasks.  But 
Asterisk aspires to higher goals with certain firms and certain 
tasks, and the generic processors on which it currently runs just 
can't hack it yet.  Echo cancellation and transcoding still are out 
of reach for Asterisk when compared with DSP-based platforms, if 
you're talking about density of channels per RU or per dollar.   Yes, 
I've done the cost comparisons with (as an example) RLX blade servers 
vs. Cisco AS5800, and the 5800 loses in cost, but not in space (it's 
smaller) and certainly not in hassle.  I can get a Cisco box up in an 
afternoon (and get physical interfaces as a bonus) but getting a 
dozen RLX boxes turned up and happy isn't something to be done 
without significant planning, standardization, administration, 
MANAGEMENT (see my post of a day or two ago) etc. which adds up.  In 
this industry, time is usually more valuable than long-term capital 
expense budget (whee!  sounds like 1998, doesn't it?)

Maybe in two years (maybe more, maybe less) the processor market will 
crest that magic cost vs. performance barrier where I'll change my 
mind, and this can be done on a smaller number of physical systems. 
After all, the processing load for dealing with human voice remains 
consistent, and processors are getting faster.  But not yet.

Today, for the truly "massive" scale, perhaps blade servers win when 
you have full time staff to do nothing but care for and feed hundreds 
upon hundreds of Asterisk servers.  Google has proven their point. 
But there are step functions here to consider, and cost is not the 
only factor that must be considered - no one solution is right for 
everything, and claiming that DSPs are dead is a bit premature at the 
carrier end of the spectrum.

The point of this is: I'd love to see an _inexpensive_ board that 
does DSP co-processing and echo cancellation, that plays well with 
Asterisk.  Getting better density (and better echo cancellation!) 
would be a big win, and if it could be less costly than the Digium 
physical interface boards of the same capacity, then that might be a 
nice combination.   I'd rather pay an additional $1200 for a card I 
can move around into various other platforms rather than pay an 
additional $1500 for the upgrade to a dual 3.0ghz system I'd need to 
handle (maybe) the same number of channels.

(before someone mis-quotes me as saying Asterisk doesn't work for 
carriers, I'd suggest you read this message over again and 
contemplate very closely the exact wordings I've used.  If English 
isn't your primary language, please be careful of the landmines I've 
laid.)

JT