how fast were drum memories?

[Ed Sharpe]

SOME OTHER  DRUM  USE  AS  I REMEMBER IT......
 FOR  STORAGE HP 2000A TIMESHARE SYSTEM  USED  AS  DRUM    AS  SYSTEMS  WERE UPGRADED AND DISCS ADDED  FOR  2000C  ETC THRU F   SOME KEPT THEIR  DRUM  AS  SWAPPING  MEDIA. AS   THE HEARD PER TRACK WAS  FASTER THAN   MOVING HEAD  FOR  USER  SPACE  AREA  SWAPPING
----  AMD -------AND  SOME OF THE GE-PAC  PROCESS CONTROL STUFF  USED  VERMONT  RESEARCH   DRUMS...  ED# 
 
In a message dated 5/10/2018 7:29:19 AM US Mountain Standard Time, cctalk at classiccmp.org writes:

 
 Drums were used as main memory in a number of early computers, and as secondary memory for a while longer. I wonder how fast real ones (actually constructed) managed to be.

What prompted this question is reading an interesting document: https://ir.cwi.nl/pub/9603 (in Dutch), "Principles of electronic calculating machines, course notes February 1948" by Prof. A. van Wijngaarden at the Mathematical Center (now CWI) in Amsterdam. It's quite a fascinating short introduction into computing technology of that era. (One comment in the intro: "The field is new. At the moment, the Eniac is the only working machine..." -- probably not quite accurate given some classified machines, but not too far wrong.)

The section on main memory describes a bunch of different technoly possibilities, one of them drum memory. He writes that a drum of 8 cm diameter (a bit over 3 inches) and "a couple of decimeters height" could hold maybe 100k bits, with a track pitch of "a few millimeters". So far so good. He goes on to suggest that such a drum might spin at 1000 revolutions per second, i.e., 60,000 rpm. That seems amazingly high. I could see it being physically possible for a drum of only 40 mm radius, but it sure doesn't sound easy. It's a good goal to strive for given that the logic, even in the days of vacuum tubes, can run at cycle times of just a couple of microseconds. As one more way to speed things up he suggests having multiple rows of read/write heads, where the addressed word would be picked up by whichever head sees it soonest. 10 rows and 60k rpm would give you 50 microseconds average access time which "even for a parallel computer would be a very attractive number". (Pages 17-18)

I'm wondering what the reality of fast drum memories looked like, and whether anyone came even close to these numbers. Also, am I right in thinking they are at least in principle achievable? I know I could run the stress numbers, but haven't done so.

 paul