Path: utzoo!mnetor!uunet!lll-winken!lll-lcc!pyramid!decwrl!spar!malcolm
From: malcolm@spar.SPAR.SLB.COM (Malcolm Slaney)
Newsgroups: comp.arch
Subject: Re: Why build TF-1 if you have its uniprocessor chips?
Message-ID: <30@spar.SPAR.SLB.COM>
Date: 2 Apr 88 00:39:01 GMT
References: <12176@brl-adm.ARPA> <1988Mar11.215238.976@utzoo.uucp> <11437@duke.cs.duke.edu> <4297@hoptoad.uucp>
Reply-To: malcolm@spar.slb.com (Malcolm Slaney)
Organization: SPAR - Schlumberger Palo Alto Research
Lines: 22
Summary: TF-1 makes good speech recognizer

In article <4297@hoptoad.uucp> gnu@hoptoad.uucp (John Gilmore) writes:
>'Twould seem that rather than build one TF-1 with 32,000 chips, you
>would do better selling 32,000 workstations (~ half of all the Suns
>in existence) that ran 16x as fast as Sun-3's.

Well, a single 50MIP processor sure sounds nice but I remember some IBM speech
people describing how they thought real time connected word natural speech
recognition would consume a good portion of the TF-1 (or whatever it is
called.)  I suspect that natural (connected word, speaker independent) 
speech recognition done cheaply will be a significantly bigger market than
Sun has ever dreamed about.

Speech recognition has always been a problem where brute force can make
for big wins.  I think the numbers Fred Jelink (sp?) presented at the DARPA 
Speech Recognition workshop were based on the current "best" algorithms
with the desired vocabulary (>5000 words) and a complete english grammar.

I'm sure that lots of other people can come up with uses for a machine with
32k processors.  This sounds to me like a Connection machine done right :-).

Cheers.

								Malcolm