23 Sep
2022
23 Sep
'22
6:41 p.m.
On Sep 23, 2022, at 2:25 PM, Chuck Guzis via cctalk
<cctalk(a)classiccmp.org> wrote:
On 9/23/22 10:52, ben via cctalk wrote:
There is of course the CDC 6000 series ECS and its successors, a bulk memory that does
block transfers to/from main memory at full memory speed.
Re 100k drum, interesting that someone tried to build that. It might actually work as a
disk, with floating heads in low pressure air rather than vacuum. But it reminds me of a
computer design course from 1948, where a discussion of memory technologies postulates a
drum memory (main memory in that era) described as "8 cm diameter and a few
decimeters long" spinning at 60k RPM to deliver an average latency of 50
microseconds. It is perhaps significant Adriaan van Wijngaarden, the author of that
document, was a mathematician rather than a mechanical engineer. :-)
Just how do the supercomputer do i/o for all that
floating numbers.
Weather maps I can see for output, but what about all that Top Secret
number crunching.
Well, consider the 1969 STAR-100; although not well documented, had a
512-bit wide, error-checked I/O channel that ran at memory speed. Neil
had various schemes for it, including a 100K RPM head-per-track drum
that ran in vacuo. I recall him mentioning that the prototype lasted
around a minute before the observation window was covered with the
remnants of the drum surface. Or consider the STAR SCROLL--a very wide tape that ran
over a
head-per-track drum. I don't recall seeing that prototype; maybe it
existed only in the mind. But we had to mention both in our responses
to RFQs.
Shades of the CDC 626, a one inch wide 14 track tape drive. That was a real product, I
think, though I never ran into one.
Another way to get high speed: one of CDC's first disk drives, the 6603, wrote several
bits in parallel. 4 bits? 12? I don't remember, but it made for a throughput spec
that was unbeaten for nearly a decade.
paul