1994-08-19 - Some facts on the Cray-3 deal

Header Data

From: pcw@access.digex.net (Peter Wayner)
To: cypherpunks@toad.com
Message Hash: 74c64c79e95425f286e4ac0cd40e529158940ae4bde64a5b0e5f21eba091a96a
Message ID: <199408192105.AA22741@access3.digex.net>
Reply To: N/A
UTC Datetime: 1994-08-19 21:05:47 UTC
Raw Date: Fri, 19 Aug 94 14:05:47 PDT

Raw message

From: pcw@access.digex.net (Peter Wayner)
Date: Fri, 19 Aug 94 14:05:47 PDT
To: cypherpunks@toad.com
Subject: Some facts on the Cray-3 deal
Message-ID: <199408192105.AA22741@access3.digex.net>
MIME-Version: 1.0
Content-Type: text/plain



<FORWARDS deleted....>

I should note that the processors on this machine sound strikingly like
the Coherent Memory Chips that I turned into DES crackers.

Also, the SRC's observation that the routing network of the SIMD machines
is expensive is an old one at Coherent. Their memory chips were only
connected with a one dimensional network and Chuck Stormon felt that
this was the only connection arrangement that scaled. This is one
of the sad limitations of packaging. In SIMD machines with multiple
processors per chip it is often impossible to quadruple the number
of processors per chip when a new fabrication process emerges. There
just aren't enough pins available. 

It turns out that the 12 dimensional routing network of the Connection
Machine was really overkill. A three dimensional one was nice on some
problems, but you paid the price in performance. I'm not surprised
that the Terasys could really run rings around the CM-2 for some 
problems that didn't need the bandwidth. It does sound interesting
if the scatter/gather hardware can do a good job with the routing.

(It is interesting that this announcement came on the same timeframe
as the obituaries for Thinking Machine. They all said stuff like 
"These guys were brilliant and the machine was great...but the market
didn't see their brilliance." It would be funny if the old Cray
design proved to be just as versatile.)

All that being said, I don't really think that this machine will be
anywhere near as powerful as the one I described. The 512,000 processors
would probably take about 300-1000 days to do a brute force DES search.
The main advantage is that the processors have 64 bits available in 
memory, not 42. That's just enough to store a complete DES (or SKIPJACK!!!)
block in place. There is no need to use the communication hardware
to go back and forth. I'll do a more complete calculation later.

I don't know whether they'll be able to add more SIMD processors with
time. 512k words of memory seems pretty small even when there are 8 bytes
to a word.  


-Peter Wayner

>Date: Fri, 19 Aug 1994 13:30:42 -0400 (EDT)
>From: Ronald G Minnich <rminnich@descartes.super.org>
>Subject: Re: SSS attachment - is that Splash II?
>To: jms <jms@central.cis.upenn.edu>
>Cc: f-troup@AURORA.CIS.UPENN.EDU
>
>The WSJ article on the Cray-3 deal involves a chip/system designed here
>called TeraSys. Maya [Minnich djf] has a forthcoming article in IEEE
>Computer about it.
>The basic idea is to embed bit-serial processors in the sram chips, one
>processor per bit in the memory row register. It's a simple processor with
>an ALU and three registers. The memory "words" can now be thought of as
>running vertically up the columns, rather than horizontally along the
>rows. Every time you fetch a row, you are in fact fetching 64 bits from 64
>different words to be operated on in the ALUs.This gives you roughly a
>100-fold increase in memory bandwidth, since you no longer mux the data
>down from 64 bits to 1 or 4 bits to bring it off-chip. The key
>contribution of the Cray (this idea due to Ken Iobst of SRC, who also was
>the architect of the chips) came from Ken's realization that the Cray vector
>scatter/gather hardware could push the bits around BETWEEN the chips as
>fast as the hardware networks on, e.g., the CM-2. Thus the Cray-3, a
>vector supercomputer, can function as a very high-performance SIMD system
>as well. You don't need to build the additional network found on so many
>SIMD machines, e.g. maspar or cm-2.  You also can have phases of a
>program, wherein it runs SIMD for a while, then vector, then SMP mode,
>etc.
>
>The TeraSys processors built here at src cost about $80K each, and for
>several problems could easily outrun our $6M CM-2.
>
>The chips are National Semi SRAMS, with mods done here by Mark Norder and
>Jennifer Schrader. Ken Iobst is the architect of the chips. The
>programming environment was basically Maya Gokhale's DBC language, which
>also runs on Splash-2, the CM-2, and clusters. Cray has reimplemented the
>chips for the Cray-3.
>
>In keeping with the NSAs new plans for SRC, this project was terminated
>at SRC last January. The systems are gradually being turned off and the
>cabinets put to other uses. No further work is occurring at SRC in this
>area.
>
>Feel free to forward this article to people who may ask you. There was
>some good work done by people here.
>
>ron
>
>rminnich@super.org     | Error message of the week:
>(301)-805-7451 or 7312 | NFS server localhost not responding still trying
>
>
>







Thread