The 360/91 and associated machines

[gah]

"William H. Blair" wrote:
(snip)

The machine originally announced as the Model 92 was, in fact,
built and delivered, but it was called a Model 95. There is a
story behind that, which most books and web sites, trying to
simplify things (I guess), get wrong.

I have both "Memories that Shaped and Industry" and
"IBM's 360 and Early 370 Systems", are those the ones that
get it wrong?

The original plan for the NPL 604 (i.e., S/360 Model 92) included
sixteen-way storage interleave, quickly discovered to be way too
expensive (and hard to build). The other hardwired machine in the
System/360 family, the Model 75, achieved only four-way storage
interleave, and then only with four huge, heavy, hot, completely-
separate 256KB storage units.) But there was a slower version of
the Model 92 in the (then-unannounced) plan: the Model 91. So the
(announced) Model 92, which was really intended to keep customers
away from the CDC 6600 (with its own problems), was essentially a
high performance (but still unannounced) Model 91 (for which only
8-way storage interleave was planned). Got it so far?

(big snip)

My understanding having read many of the IBM and IBM related
documents is that the 91 had 16 way 750ns core. The 95
had (past tense, as far as I know none exist anymore)
magnetic thin film memory. The 91 goal was one operation
per 60ns cycle, so with 750ns core it would need 16 way
to have any chance. Other than that, I don't think I can
disagree with any that you said. I did much of my early
programming, including my first assembly programming, on
a 91.

Also, it was my understanding that a 360/85 was delivered
from when a 91 was ordered until it was ready. Stories
are that with the cache the otherwise slower 85 came
close to the speed of the 91 on most real problems.

-- glen

[Peter]

Post by gah
My understanding having read many of the IBM and IBM related
documents is that the 91 had 16 way 750ns core.

Then they were using modified M65/M75 memory, which was 750 nsec
(interleaved four ways on those models).

Cache would have helped the M85 a lot vs. a non-cached M91.

[William H. Blair]

[part 2, continued from part 1]

Post by gah
My understanding having read many of the IBM and IBM related
documents is that the 91 had 16 way 750ns core.

The Model 91 had 16-way interleave Model 2395-1 or 2395-2 (or
both) core memory units. Each one was independently accessed,
although depending on where it was attached (on the right or
the left side of the 91 processor frames cross), one was just
a tad slower than the other (not the core memory itself, but
the total actual throughput, due to differences between the
previously-mentioned MSCE and PSCE). In some sense, you could
say that a Model 91 had 32-way storage interleave, and I know
of at least one IBM document whose author claims exactly that.
If a Model 91 were ever actually able to have up to 32 memory
accesses pending at one time, it could have, perhaps, reduced
latency. But my point is that it would have been pretty hard
for a Model 91 processor to ever need, much less try, to get
any more than 16 memory accesses started at the same time.
I/O could not have driven it up very much, either. The only
really fast I/O device then attachable was a 2305 FHF, which
transferred data at 3 MB/sec. And, you could only have one.
Thus, because of the basic 2395-x storage access time vs.
the Model 91 cycle time, the machine could not really use
much more interleave than 16-way. Actually having 32-way
storage interleaving might have helped a Model 91 perform
just as well with 4 MB or 6 MB of total storage as it did
with just 2 MB "on the left," but I would bet that a 91's
actual performance with 2 MB on the left and 2 MB on the
right was in fact nearly indistinguishable from one that
had only 4 MB on the right side of the CPU (a configuration
which could have only 16-way storage interleave, at most).

While the Model 91 cycle time was 60 nanoseconds, the storage
cycle time was set to 780 ns (to match the 91's 60 ns cycle
time multiplied by 13). Now the cycle time of the storage
core arrays themselves were actually 750 ns. But the minimum
total storage access time was either 600 or 900 ns, depending
on which "side" (which SCE) of the 91 it was attached to.

There were four models of the 360/91:

Sales Total MSCE PSCE
Model Bytes High Performance Extended
------ ----- ---------------- --------
2091K 2 MB One 2395 Model 1 None
2091KK 4 MB One 2395 Model 1 One 2395 Model 1
2091KL 6 MB One 2395 Model 1 One 2395 Model 2
2091L 4 MB None One 2395 Model 2

2395 Model 1 was interleaved 16 ways with a capacity of 2 MB.
2395 Model 2 was interleaved 16 ways with a capacity of 4 MB.

Extended Main Storage had a longer total storage access time
than High Performance Main Storage, even though the actual
ferrite core storage arrays were identical, because it was
just slower. I can only speculate why this was the case but
I think it must have had something to do with the technical
requirements of supporting 4 MB total, and the longer wire
lengths that might have resulted. As I said, only speculation.

A 6 MB 2091KL 360/91 with one 2395 model 1 and one 2395 model 2
was physically huge: on plan diagrams it takes the same amount
of space as approximately fifteen (15) z10 processor frames!

But how did they achieve that 16-way interleave on each 2395?
Each 256 KB (four 64 KB core arrays) was two-way interleaved.
So, one 2395 frame (512 MB) was four-way interleaved. A 2395-1
had four 512 MB frames, for sixteen-way interleave. A 2395-2
had eight 512 MB frames, but there was only circuitry for 16-
way interleave in the 2395-2 PSCE.

Post by gah
The 95 had (past tense, as far as I know none exist anymore)
magnetic thin film memory.

Yes. It, too, was 16-way interleaved, because it had 16 "units."
But those 16 units totaled only 1 MB. If it could have more, no
other Model 95s were ever built that had any more than sixteen
64 KB units, for a total of 1,024 KB, not including the 4 MB of
regular Model 91 memory, which, you will note, was IBM Model 2395
memory, numbered that because it was, in fact, the original core
memory unit that IBM had planned to use for the 92 which became
the 95 when they discovered it was not going to be fast enough.
That is just one little subtle fact that illustrates that my
version of the sequence of events is the correct one. The 4 MB
of core storage on the Model 95 was exactly the same as the only
kind of storage (core) that you could get on the Model 91: 2395.

I do not know which model 2395 boxes were actually used on the
only two Model 95s that IBM did build, but I was once told that
one came with one 2395-1 main storage unit and the other came
with a 2395-2 main storage unit. If that is true, then one Model
95 had 3 MB total and the other 5 MB total.

Post by gah
The 91 goal was one operation per 60ns cycle, so with
750ns core it would need 16 way to have any chance.

Exactly. But as I said, they slowed it down to 780 ns to be
an exact multiple, 13, of 60 ns.

Post by gah
Other than that, I don't think I can disagree with any[thing]
that you said. I did much of my early programming, including
my first assembly programming, on a 91.

I'm jealous. My first big, fast machine was merely a Model 75.

Post by gah
Also, it was my understanding that a 360/85 was delivered
from when a 91 was ordered until it was ready. Stories
are that with the cache the otherwise slower 85 came
close to the speed of the 91 on most real problems.

That's about right. On scientific workloads a 360/195 was
about three times as fast as a 360/85, even though its
circuits were only twice as fast. The improvement was due
to the fact that a 350/195 was just a 360/91 implemented
using monolithic integrated circuits, like the ones later
used with System/370. On "real" problems (programs that
were not FORTRAN scientific matrix / vector manipulation
with tight inner loops) a 360/91 would only make around 3
MIPS. The cache on the slower-to-begin-with Model 85 was
able to pull it up into the Model 91's "commercial work"
(i.e., "real" problems) performance range -- well, almost.
IBM's super-fast 360/91 with a cache (which they called a
360/195) pulled out ahead of the lot, mostly because of
improvements to the cache over a Model 85, but some other
improvements as well.

Post by gah
Then they were using modified M65/M75 memory, which
was 750 nsec (interleaved four ways on those models).

The Model 91 used "2395-1" or "2395-2" memory. Internally
it was the same ferrite core storage arrays used in 2365-3
memory units for Models 65 and 75; externally these frames
were huge. A 2395-1 was a U-shaped addition to one end of
the 2091, whose physical size was equivalent to six 2365s.
A 2395-2 was an H shaped addition (two Us put back to back),
whose physical size was equivalent to ten 2365s. The bottom
of the U (and the cross of the H) held uninteresting stuff,
I think, so it was the uprights of the U and the H which
held 1 MB each. Since the size of one U upright was nearly
the same as two 2365-3 units, but held four times as much
core storage (and I know how packed a 2365-3 was), I think
that the explanation for the discrepancy is that the power
supply and other hot, heavy stuff present in a 2365-3 frame
was not located in the uprights of a 2395 frame, but in the
cross of the H or the bottom of the U, thus freeing enough
space to fit eight 64 KB cubes of fast core memory in each
(physical) frame, instead of four, which is all that would
fit in a 2365-3 frame, as can be seen at:
http://www.staff.ncl.ac.uk/roger.broughton/museum/corestore/ram64k.htm.

--
WB

[William H. Blair]

[part 1 of 2 parts]

Post by gah
I have both "Memories that Shaped and Industry" and
"IBM's 360 and Early 370 Systems", are those the ones
that get it wrong?

I have only the latter, but I have not read it in years. What
I do remember is that virtually everywhere I go on the web,
and in most books I have actually read, they get the very fine
details (mostly, the time sequence) of what actually happened
wrong. I'd have to check my copies to answer your question.
But it's really not that important, obviously. I just hate it
when folks mindlessly pick up something they read online, and
accept it as authoritative because it sounds that way. If I
were writing a book for a reputable publisher, their fact
checkers would have a tiff with me because most places they
could go to check facts (that I have found) just have them
plain wrong to begin with. So, I'd probably get "corrected."
At least what I know is now being archived, here if nowhere
else.

For example, look at this web site:
http://ed-thelen.org/comp-hist/vs-ibm-360-91.html
Unless the first part was talking about a specific machine that
was installed at some specific customer site, this page is very
confused about the maximum amount of memory that a Model 91 can
have. It first says it's 4 MB, and then just below it says 6 MB.
The first two sources they mention are books that I have, but I
have not yet checked them for accuracy. But I bet they get some
of the details wrong. The guys that wrote those books were not
working at this level of detail on a daily basis (if ever), so
I would not even expect them to recognize mistakes of this sort.
Quite understandable, in fact. But the publisher's fact checkers
should have made them prove it, since these books were actually
little more than collections of facts (not much plot, anyway).
Obviously, they didn't do that. Go figure.

Even this IBM web site:
http://www-03.ibm.com/ibm/history/exhibits/mainframe/mainframe_FS360B.html
indicates that a 360/95 can have from 1 MB to 6 MB of 780 ns
storage. Along with the 1 MB of thin-film storage every Model
95 had (yeah, all two of them!) that would add up to a max of
7 MB, not 5 MB as most sources indicate. More on that issue,
later (below). At this point (40 years later, almost exactly)
it is probably like counting angels on the head of a pin, and
who knows what actual differences there were between the ten
or twenty (more on THAT later, too) Model 91s and the two 95s.
The two 95s, for example, could actually have had provision
for attaching 6 MB of additional core storage, but that would
have required a third "Storage Control Element." A Model 91
had two storage control elements, one for the left side of
the box, and one for the right. The "Main Storage Control
Element" was connected to the "High Performance Main Storage"
(HPMS) unit, which was (more on this, later) a 2395 Model 1.
The "Peripheral Storage Control Element" was connected to the
"Extended Main Storage" (EMS) unit, which was either a 2395
Model 1 or 2. Now, I VERY STRONGLY SUSPECT that in the only
two real Model 95 that IBM built the thin-film monolithic
memory was attached to the Main Storage Control Element. The
fact that Models 91 and 95 had two distinct Storage Control
Elements in the first place was due to the fact that what
became the 95 had to support BOTH the 1 MB thin-film stuff
as well as ordinary core memory. Thus, although no Functional
Characteristics manuals exist for the Model 95, I'll bet just
about anyone that the Main Storage Control Element, which had
higher performance than the Peripheral Storage Control Element
in the first place, was the attachment point for, and was
exclusively dedicated to, the thin-film memory. This left the
slower Peripheral Storage Control Element available for the
4 MB 2395-2 core storage "H box." All of this, therefore, just
means that, for obvious technical reasons, the maximum storage
capacity of any Model 95 (as then designed) was 1 MB + 4 MB =
5 MB, not 7 MB. (If it had been anything else, it would not
have been a 95, but a 91.) So, I don't have to know very much
about two machines that were special order-built to be able
to know this fact. All I have to know is that the 95 was a 91
and how all 91s were constructed internally. But what I still
do not know is if the two 95s had one 2 MB 2395-1 core storage
unit or one 4 MB 2395-2 core storage unit actually installed.
I believe that I remember being told that one had 3 MB total,
and another had 5 MB total. If so, one came with a 2395-1 and
the other came with a 2395-2. But all my contacts at NASA in
NYC are long lost, so I doubt that we will ever actually know.

Here is another IBM web page:
http://www-03.ibm.com/ibm/history/exhibits/mainframe/mainframe_2423PH2095.ht
ml
which states:

"The Model 95 -- two of which were installed in 1968 at NASA's
Goddard Space Flight Center in Greenbelt, Md. -- were used to
solve space exploration problems requiring unusually high
computation speeds."

But then here is another site:
http://www.quadibloc.com/comp/pan05.htm
which states:
"Only two IBM 360/95 computers were made, both for NASA. One
went to the Goddard Space Flight Center, in Greenbelt, [MD],
and the other went to the Goddard Institute for Space Studies,
part of Columbia University, in New York City."

Guess which one is right? Not the first (IBM) one! But how do
I know? Well, let's take a look at another IBM web site:
ftp://ftp.software.ibm.com/eserver/zseries/misc/bookoffer/download/360revolu
tion_040704.pdf
Here we can read:

"One model of the System/360, the 95, was built
especially for NASA, and only two units were made.
One went to Goddard Space Flight Center in
Maryland; the other to the Goddard Institute for
Space Studies on upper Broadway in Manhattan.
David B. Soll, who is currently a senior technical staff
member for IBM's Business Continuity and Recovery
Services, was a NASA contractor at GISS in New
York in the 1970s and 1980s. "The model 95 was a
supercomputer of its day," he said recently. "With it,
we carried out research in Meteorology and
Climatology, developed instrumentation for a Landsat
satellite to measure agricultural data, and instruments
that traveled to Venus and Jupiter aboard the
Pioneer Venus and Galileo spacecraft." "

Now that sounds a little more authoritative. But if you look
at the picture on page 38 of that document, you will see that
it is identical to one on many other web pages that Google
will easily reveal to you where it is described as a 360/91!
Here are the most egregious two examples:
http://www-03.ibm.com/ibm/history/exhibits/mainframe/mainframe_PP2091.html
http://www-03.ibm.com/ibm/history/exhibits/mainframe/mainframe_2423PH2091.ht
ml

My question now is: Is that picture of a Model 91 or of
a Model 95? And where was it taken, exactly? Who are the
people? (I bet they are IBM, or IBM contractors, not NASA
folks.)

But then this web page:
http://www-03.ibm.com/ibm/history/exhibits/mainframe/mainframe_PP2095.html
seems to say (and we don't even have to read too much between
the lines) that both Model 95 went to Greenbelt, MD:

"Formal acceptance of two, new super-speed computers -- IBM
System/360 Model 95s -- by NASA's Goddard Space Flight Center
was announced here (Washington, D.C.) today by International
Business Machines Corporation.

"The two computers are the first and only ones in IBM's Model
90 series equipped with ultra-high-speed thin-film memories.
Over a million characters (bytes) of information are stored
in each on magnetic "spots" four millionths of an inch thick.

"Both of NASA's Model 95s are handling space exploration problems
which require unusually high computation speeds.

"One Model 95 is the primary data processing facility for the
Center's Tracking and Data Systems Directorate. It provides
additional computing support to the Project, Technology and
Systems Reliability Directorates.

"The second Model 95 is being used by astrophysicists at NASA
to create massive mathematical models of the universe. The
boost in computing capability is enabling them to simulate
the evolution of galaxies, stars and planets to a degree
never before possible."

But they really didn't say exactly where the second one went,
now did they? I happened to know a programmer who worked at
NASA at Columbia University. He told me about the Model 95
they had there long before this most recent IBM document was
researched and appeared. He was in a position to know. And I
trust him. But that still does not answer the question about
which machine is actually in which picture. I may never know
and nobody will unless one of the original participants ever
shows up (or the photographer). The 95 and the 91 console
were identical, so we can't tell them apart even if we did
have a good sharp photograph to analyze (which we don't).

NASA had one 360/95 at NASA (per se), but they also had some
360/91s, one of them being particularly famous, and that may
have confused the poor (young) folks at IBM who researched
and prepared material both for ex-IBM executives who wanted
to write books as well as present-day IBM history web sites.

Others are confused about the shipping dates and sequence of
Models 91 and 95, not just their basic, raw specifications.

Many disagree on the number of 360/91s that IBM built or sold.
I have read and heard it authoritatively stated that the number
was 10, 11, 12, 14, 15, or 20. The above web site states:

"According to IBM's 360 and Early 370 Systems by Pugh, Johnson,
and Palmer, IBM produced fourteen Model 91 computers, four of
which it kept for internal use."

Oh, my! What should we believe? I think, but cannot prove, that
the actual number was 15, and IBM may indeed have kept four of
them. That would be consistent with other reports that IBM SOLD
11 360/91s. That figure (the number actually SOLD) would figure
(pun intended) later in IBM's legal briar patch over the CDC 6600
reaction issue. So, while I cannot PROVE that IBM built 15, I
have a pretty good idea that's the truth, and many other facts
are consistent with that figure. Still, most web sites (and, I
suspect, most books) get the number wrong. Again, not important
except that if the point of some document is to get facts right,
shouldn't they be right or at least advertised as "reported" or
"unconfirmed"?


[continued in part 2]

[Peter J Farley III]

Post by William H. Blair
[part 1 of 2 parts]

<Snipped>

Post by William H. Blair
http://www.quadibloc.com/comp/pan05.htm
"Only two IBM 360/95 computers were made, both for NASA. One
went to the Goddard Space Flight Center, in Greenbelt, [MD],
and the other went to the Goddard Institute for Space Studies,
part of Columbia University, in New York City."
Guess which one is right? Not the first (IBM) one!

And I can personally testify that you are correct about this fact.
Back in the day (ISTR it was 1970 or perhaps 1971, but the memory is
somewhat fuzzy at this remove) I interviewed for a TSO support job
at Columbia University (the job ultimately went to an insider at the
university). Part of the half-day-long interview process was a tour
of the computer center, which had a 360/91 being fed and watered by
a 360/75. I was told at the time that ASP (not HASP) was being used
by the 360/75 to feed work to the 360/91 and retrieve and print the
results.

It was quite an impressive sight to my then-young eyes. I always
regretted that I didn't get that job.

Peter

[Peter]

Post by Peter J Farley III
I was told at the time that ASP (not HASP) was being used
by the 360/75 to feed work to the 360/91 and retrieve and print the
results.

Originally, the larger models were not intended for unit record
equipment, so multiplexer channels (later became the 2870) weren't
available.

U/R was supposed to be transferred to tape on a smaller machine,
which was read by the larger machine's selector channel (2860).

HASP and ASP changed most of that.

And, selector subchannels were added to the multiplexer channel.

[jeffsavit]

Post by William H. Blair
If I
were writing a book for a reputable publisher, their fact
checkers would have a tiff with me because most places they
could go to check facts (that I have found) just have them
plain wrong to begin with. So, I'd probably get "corrected."

Fact checkers at the publisher? Oh dear me, that's a fantasy. When I
did "IBM Mainframes" (ISBN-13: 978-0070506916) and other related books
I had to fight off a copy editor, but no fact checking was ever done.

In fact, when it was time for me to review galley proofs for "VM/CMS:
Concepts and Facilities", I had a copy editor to "helped" me by
replacing every reference to "DOS/VS" or "DOS/VSE" with "MS-DOS", and
every reference to "OS/360" with "OS/2". Boy, did I ever have a
conniption. I guess they never heard of mainframe operating systems. I
had to have an emergency call with the publisher to get all those
changes removed. So, I wouldn't count on any fact checking!

regards, Jeff

[William H. Blair]

I received the email, forwarded below, from a friend
who has given his permission to forward its contents
to this list.

| From: "David B. Soll"
| Sent: Monday, April 14, 2008 5:14 PM
| To: William H. Blair
| Subject: 360/95 Reminiscences
|
| I worked at the Goddard Institute for Space Studies (GISS)
| in NY, from 1974 to 1983. I was hands-on with the 360/95
| installed there - as you correctly indicated, one of two.
|
| My recollection (now 30 years old) is:
| 4 MB of 780 ns core memory, banked 16-way
| 1 MB of 60 ns READ/Refresh thin film memory
| I heard (but never verified) that there was an enhancement
| to the Floating Point Divide.
|
| The console was almost identical to that of the 360/91,
| with one exception. Included on the console was something
| called the M120 Bypass control, which permitted the machine
| to bypass use of the thin film memory and address the first
| BOM of core memory with real address zero. This allowed for
| the repair and/or diagnosis of problems with the thin film
| array.
|
| In addition, the OS used was an MFT-like OS designed and
| built at GISS called "SSS", Scientific Supervisory System,
| author Paul B. Schneck, then Director of Computing Faculties
| at GISS.
|
| My recollection is that when Floating point operations were
| properly organized to use the 8 double word instruction stack,
| we actually achieved 5 MIPS. Because the 360/95 did not have
| decimal instructions, but rather used a software simulator,
| we never used it for any commercial applications.
|
| I hope this helps..
|
| Regards,
|
| David B. Soll

I have David's email address, but have removed it from the
forwarded message, above, for obvious reasons, despite the
fact that David said I was free to publish it. Anyone who
wants to confirm this report is welcome to ask me for his
email address.

Now, if I could only turn up some person who actually worked
with the only other 360/95 at the Goddard Space Flight Center
in Greenbelt, MD, and remembers what its memory configuration
was, I could authoritatively correct the record everywhere.

I have said that I was told that one had 3 MB and the other
had 5 MB. We now know, courtesy of David Soll, that the one in
NYC had 5 MB, so the one in Greenbelt must have been the one
that had 3 MB -- IF my (old) information is correct. But I do
not really know that it is. I always thought that the 360/95
in Greenbelt was the one that had 5 MB, so it may be the case
that my old information is incorrect; if so, then both 360/95s
had a single 2395-2 main [core] storage unit.

In a subsequent email message, David indicates that he has
some old manuals for the 360/95 and the 360/195. I will be
acquiring those manual to scan them into PDFs so that they
can be made available to others.

Gosh, dawg ... I do love pack rats like ourselves!

--
WB

[Mike Schwab]

What was the relationship between the New York and Greenbelt? Was
Greenbelt the head office? Would the report of a successful NY
installation be reported to Greenbelt instead of directly to IBM?

[Peter]

Post by Mike Schwab
What was the relationship between the New York and Greenbelt? Was
Greenbelt the head office? Would the report of a successful NY
installation be reported to Greenbelt instead of directly to IBM?

IIRC, the very first Amdahl 470V/6 (S/N 00001) replaced (or
augmented) the machine in New York City.

[Gregg C Levine]

Hello!
As I understand it, Mike, it is simply a matter of circumstances. IBM has
its headquarters here in the city since before the first computers as we
understand them went active.

A large number of them, went to work in various universities, amongst them
was of course Columbia. In fact there's a good story behind it all someplace
on the IBM website on their historical section.

Greenbelt is of course the home to the Goddard Space Flight Center, and is
reasonably closer to the Federal Systems Division which is in VA.

I haven't worked out the reasons either but it is also a strange series of
events regarding the ones living in Boston and Cambridge.

Oh and thank you Peter for that bit of trivia.
--
Gregg C Levine hansolofalcon-XfrvlLN1Pqtfpb/***@public.gmane.org
"The Force will be with you always." Obi-Wan Kenobi
 

-----Original Message-----

On

Behalf Of Peter
Sent: Tuesday, April 15, 2008 6:36 PM
Subject: Re: [hercules-390] The 360/95 at NASA Goddard Institute for Space

Studies in

NYC

Post by Mike Schwab
What was the relationship between the New York and Greenbelt? Was
Greenbelt the head office? Would the report of a successful NY
installation be reported to Greenbelt instead of directly to IBM?

IIRC, the very first Amdahl 470V/6 (S/N 00001) replaced (or
augmented) the machine in New York City.

[William H. Blair]

Post by Mike Schwab
What was the relationship between the New York and Greenbelt?

The same as today, I suspect. Not much has changed.

See http://www.giss.nasa.gov/ and, of course (now):
http://www.nasa.gov/centers/goddard/home/index.html

Post by Mike Schwab
Would the report of a successful NY installation be
reported to Greenbelt instead of directly ... ?

No. It wasn't at the time. The problem is that there
are very few tracks left of that event on the World
Wide Web, which didn't come along en masse until ~25
years later. Those of us around back then are trying
to get old documents scanned and the machine-readable
history archived -- before we lose interest (or just
any more of our memory).

The 360/95 was a big thing, although it later turned
out that the 360/91 was just as big a thing, because
the 1 MB of fast thin-film memory on the Model 95 was
just not able to help that much. Why? Well, in order
to get the 5 (floating point, et al.) MIPS out of the
box, you had to construct a program that would turn
the yellow "loop light" on the huge operator's panel.
When this happened (this required only ordinary effort,
and you can find stories about this even today using
Google; e.g.: www.columbia.edu/acis/history/ and
www.columbia.edu/acis/history/elliott-frank.html#loopmode)
then the entire sequence of instructions were held in
the I unit pipeline, and no more references to core
or thin-film memory were made, anyway (that is, for
instruction fetch). Given that the 2395 core storage
units were sufficiently interleaved to provide data for
and store results from a new instruction being executed
every cycle -- it would then run at top speed producing
a new result to be stored every cycle -- it did not
matter that the core storage was relatively slow 780 ns
memory. The 16-way interleave permitted the fetching
and storing of far more data than the box would need
or could produce when running at absolute top speed,
in "loop mode." Hence, for what the Model 95 had been
designed to do, the thin-film memory was not needed
at all. It MAY have helped a little for what the box
was NOT designed to do, but it was so darn slow when
running a non-loop mode-achieving program (basically,
about 1.5 to 2.0 times the speed of a Model 75, which
was also a hardwired CPU) that the faster monolithic
memory could not have contributed that much (esp. when
compared to the 4-way interleave of the 2365-3 750 ns
memory boxes on the Model 75). In other words, even a
(slower) 16-way interleaved 2395-2 core storage unit
on the Model 91 was all that a machine like a Model 91
needed in the first place. When running "fast," it did
not USE any fast memory, and when running slowly, it
simply did not NEED that fast, thin-film memory (most
of the time).

Of course, most of the world did not know at the time
that the 360/95 was relatively sluggish when running
non-scientific programs, even worse if you executed
any decimal instructions (other than ED and EDMK). The
360/95 was supposed to be a really fast supercomputer!
Well, it was, but ... (insert reams of discussion).

Since IBM lost money on all of the Model 9x machines,
they would have milked it for all it was worth, since
they had already stuck their foot in their mouth by
even announcing it/them 3-4 years early. Any good
publicity they got out of it was evidence that the
boxes were real. And, they did meet their announced
performance goal. In the end, what really irritated
CDC was IBM's "disruptive" (to CDC's business, that
is) announcement of what turned out, in fact, to be
a loss-leader (so to speak). By the time the very
first Model 91 shipped (and that was announced), IBM
had discontinued the machine, cutting it from the 360
lineup, and announced that the program was really just
a research project, all goals had been achieved, and
no more orders would be accepted. Of course, what they
did not say is that one of those goals, in fact, was
to scare potential customers away from the CDC 6600.

--
WB

[Gerhard Postpischil]

Post by William H. Blair

Post by Mike Schwab
What was the relationship between the New York and Greenbelt?

The same as today, I suspect. Not much has changed.
http://www.nasa.gov/centers/goddard/home/index.html

I worked at GISS from 1963 to 1965, and from my perspective a
lot has changed. Data center support is supplied by an outside
contractor, back then it was my employer, Computer Applications,
Inc. (not one of the ones found by google.com), now it's SGT and
Sigma Space. Back then the effort was spread over three
buildings, not counting the Columbia academics. We had a
7094-II, with a Stromberg-Carlson 4020 plotter and a 1090
display; a few months before I left, we got a 360/65 in a brand
new computer room in what is now the only building. Every week
we got several cases of tapes, with data from the Tiros and
Nimbus satellites. I doubt they transmit via tape these days.

And of course with Hercules I have more processing power on my
laptop than the 91 or 95.


Gerhard Postpischil
Bradford, VT