Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!cs.utexas.edu!uunet!super!mjt
From: mjt@super.ORG (Michael J. Tighe)
Newsgroups: comp.arch
Subject: Re: Control Data Closes ETA Systems, Inc.
Summary: It didn't cut it
Keywords: supercomputers ETA-10
Message-ID: <8278@super.ORG>
Date: 21 Apr 89 03:31:47 GMT
References: <492@hydra.gatech.EDU> <16626@oberon.USC.EDU>
Sender: news@super.ORG
Reply-To: mjt@super.UUCP (Michael J. Tighe)
Followup-To: comp.arch
Distribution: usa
Organization: Supercomputing Research Center, Lanham, Md.
Lines: 55

In article <16626@oberon.USC.EDU> gshippen@pollux.usc.edu (Gregory Shippen) writes:
>Now that I had my politicoeconomic say, what does everybody think about this
>action?  Is this just a matter of bad management/sales as the HEP apparently
>suffered from or was the ETA machine just another mediocre machine that did
>not stand a chance at making ETA profitable any time soon as CDC claims?
>My understanding is that while the ETA machines architecture was nothing new
>(just another CYBER xxx right?) the engineering and packaging was a legitimate
>progression of the state of the art.  Is this so?  (Eugene do you still read
>comp.arch?).

The ETA machine had many problems, both hardware and software. The
project probably should have been cancelled years ago. It was over
budget and behind schedule.
 
The architecture was nothing really unusual, as far as I can recall.
The chips were made of CMOS, whereas the Cray Y-MP is ECL. Some
problems that I noticed included the following:
 
The air-cooled models were very loud. You could not be in the room
with one without suffering some hearing loss. The ventilation system
blew air out the side hitting you in the face if you walked by. The
CPU's and memory were deep inside the cabinet. This increased the time
and complexity to replace one (which was all too often; there were
many failures) when there was a failure.
 
For the LN models, you had to wait hours for the temperature to rise
before you could remove them. I also heard that the some of the CPU
boards on the LN models started to crack after long exposure to the
LN.
 
On the multiple CPU models (at least the one I used), the CPU's were
distinct nodes. There was no multiprocessing like on a Cray X-MP or
Convex C2xx. In fact, the CPU's had different host names and were
connected by network. Not exactly tightly coupled.
 
As for software, ETA used regular System V. That's nice for some, but
I like the BSD enhancements and so do a lot of others. (See this
month's "Communications of the ACM" interview with Steve Job's; he
almost made this same mistake with his NeXT machine, fortunately he
talked about software with some of his potential customers first and
was put back on track. If ETA had done the same...)
 
Compilers and editors were also lacking. The Fortran compiler could
vectorize, but it could not parallelize code. The C compiler could do
neither.
 
This is just a short list. I am sure others can add to it, but the
bottom line was the machine just didn't cut it. I didn't mean this to
be so negative, but that's the way it is.

-- 
-------------
Michael Tighe
internet: mjt@super.org
   uunet: ...!uunet!super!mjt