25 Oct
1999
25 Oct
'99
12:02 p.m.
On Mon, 25 Oct 1999, Carlos Murillo-Sanchez wrote:
Mike: you wanted numbers.
The 11/780 is one machine on which I do have a few recently acquired
docs. In page 2-2 of the 1977-78 VAX 11/780 architecture handbook there
is a nice figure of the cpu and different interconnects. Basically, the
cpu, main memory and I/O subsystems were tied to the Synchronous
Backplane Interconnect (SBI), 13.3 MB/s, cycling at 200ns.
Great! Numbers! Thanks, Carlos. OK, SBI = 106 Mbits/sec.
For comparison, the 486's Main Bus runs at 848 Mbits/sec [1]. That's 8 times
as fast as the 11/780 SBI bus.
The console
was attached to the CPU directly thorugh a separate channel. Any device
could request control of the SBI and control was granted using a priority
scheme. The I/O subsystem could hold one unibus adaptor (1.5 MB/s) and
up to four massbus adaptors (2MB/s each).
Unibus = 12 Mb/s (Coincidentally, USB is -also- 12 Mb/s ...)
Massbuss = 16 Mb/s each.
(For a rough comparison, the 64-bit 66 MHz PCI bus is over 4 Gbits/sec; so
that particular modern PC bus has 57 times more bandwidth than a Unibus
and four Massbus adaptors, combined!)
The cpu had 8KB of two-way
set-associative write-through cache (and the instruction set allowed for
generation of very compact code). The cache monitored I/O <=main memory
transfers and updated accordingly.
The 486 also has 8K of cache, and is approximately 30 times as fast as the
11/780 running the Dhrystone 2.1 benchmark.
Given this architecture, I think that there is reason
to assume that an
11/780 might service 100 users over serial lines faster than any 486-66
PC . Just think of the interrupt service overhead of 100 users typing
simultaneously, even with 16 byte FIFO buffers.
Well, let's see. The ISA16 bus runs at 8.33 MHz, 4 clocks, 16 bits per
cycle. That's only 33.3 Mb/s, comparable to a Unibus plus a Massbus.
(And, of course, most 486dx2/66 machines used at least an EISA bus [2],
running 32 bits at 8.33 MHz x two clock edges = 533 Mbits/sec.)
So if we are able to plug multiport intelligent serial cards into our EISA
bus, and to use a standard EISA SCSI controller, the numbers say that the
dx2/66 would crush the 11/780 by a factor of at least 5 (which is
533 Mb/s / 106 Mb/s) or perhaps as much as 30, depending on the task.
By the way, this little book is great reading... It
really puts things
in perspective to see the great 11/780 implementation 22 years ago...
I remember the book well; I wish I still had my copy!
Carlos.
Thanks again, -Mike
[1] Intel i486 Microprocessor Manual, Page 1. Doc #240440-002 November 1989
[2] Indispensable PC Hardware Book, 3rd edition, p. 552