[COFF] GE-635 GECOS?

[Clem Cole via COFF]

First, a side note.  Seymour felt the same way WRT to VM.   The truth is,
VM lets the OS and HW *manage the program overlays automatically,* so the
programmer doesn't have to think about it.  For generations of programmers
who have never used overlays and thunks, it seems strange not to have a VM.
But supporting VMs takes a lot of resources (in both the HW and the OS),
and to many people like Gene and Seymour, it seems like sloppy programming
to rely on them.  Remember, in 1965 dollars, the price per bit was $2.52
($25.26 in 2026), whereas modern DRAM is $0.000000007 per bit.  Also, a
Model 65 installation was often 512 KB or sometimes 1 MB.   Wikipedia
suggests that configurations exceeding 1 MB were considered "unusual"
during the mid-to-late 1960s.  CMU's 67 (that I used to program) had 4M,
and I think many university sites like Stanford, Michigan, Cornell, and
Princeton were likely similar.  And to scale the size, a 1M core box was
multiple cabinets and about the size of the later DEC 11/780 cabinet.

WRT the S/360 architecture, my friend Russ Robelen led the Model 50
hardware team (and was also the guy who convinced Amdahl and Brooks to use
microcode for the S/360 project).  Some years ago, Russ told me the story
of how bytes and words were defined for the S/360. The core of what Paul
says is true, but the full facts are even more interesting, and I think it
makes a great story.

While we think of Brooks for his impact on the software, he was the lead of
everything on the project, both hardware and software. Gene led the
hardware team.  Gene made it clear that he felt anything larger than a
6-bit byte and 24 bits for words and addresses was "a waste of hardware"
and would add unreasonable cost and complexity for no real reason.  As Paul
mentioned, Brooks believed that bytes and words needed to be powers of two
"because anything else is just too difficult to program."  From what I
understand, Gene was not handling that choice well.

Russ says, every time Gene brought it up, he got tossed out of Brooks'
office and told not to come back.  Again, as Paul reported, Fred did relent
on the address being 24 bits, but told Gene, "to make damned sure you store
them a full word."   So, except for the Model 67 [which was a Model 65 plus
the VM hardware called "DAT box" [Data Address and Translation], plus some
new microcode, all had 8-bit bytes, 16-bit ½ words, and 32-bit full
words].  But because the 24-bit addresses were stored and primarily passed
as 32-bit words, the 67 could run code targeted for the other models with
straightforward changes [67 had a couple of new user space
instructions - Branch
and Store (BAS), Branch and Store Register (BASR)†, and some new privileged
operations specific to the data translation hardware.]
Years later, Gordon Bell would state that the #1 issue limiting the
longevity of an ISA was too few address bits.  IBM started System 370 with
the same 24 bits, but by the 370-XA, it finally made everything 32 bits.
That architecture survived from 1964-1978 as S/360, 1970-1990 as S/370, and
1990-2000 as S/390.   It was not until the zSeries 900 that IBM broke
stride and introduced a 64-bit variant, but they do have a level of
compatibility that allows 60-year-old binaries to continue running.


† Have spent much of my youth moving OS/360 code to TSS.  Besides things
like different OS Service interfaces and system control blocks, the ugliest
thing you had to deal with was how programming used the "wasted" high-order
byte when manipulating or storing an address.  Because TSS/360 and MTS/360
used 32-bit virtual addresses, some programs that performed manual address
arithmetic or "dirty" pointer tricks (like using the high-order byte of a
32-bit register for flags) often broke because the system might need to use
those bits for extended addressing.

On Sat, Apr 4, 2026 at 12:35 PM Paul Winalski via COFF <coff at tuhs.org>
wrote:

> On Sat, Apr 4, 2026 at 9:42 AM Douglas McIlroy via COFF <coff at tuhs.org>
> wrote:
>
> > IBM had been ruled out over the issue of virtual memory. GE was happy
> > to offer a VM variant of the 635. In contrast, IBM's chief architect,
> > Gene Amdahl, adamantly maintained that VM was unacceptably
> > inefficient.
> >
>
> Gene Amdahl was a notorious opponent of the concept of virtual memory.  He
> famously stated that virtual memory merely magnified the need for real
> memory.
>
> Amdahl also had big fights with Fred Brooks over the word size for S/360.
> Brooks insisted that the machine word size be a power of two.  Amdahl
> favored a 24-bit design.  They eventually compromised:  the word size,
> registers, and arithmetic operations on S/360 would be 32-bit but the
> addressing would be 24-bit (16 MB).  Given that the biggest main storage
> ever shipped on a S/360 was 8 MB,  24-bit address circuitry was more than
> adequate and reduced the manufacturing costs for the machines
> significantly.
>
> -Paul W.
>