Section 2
The SDS 910-9300 series, a planned family
The Scientific Data System 900-9000 series consists of the SDS 910, 920, 925, 930, 940, 945, and 9300 computers. The series includes capabilities and features found in most 24-bit machines. The design implementation is among the best for 24-bit machines, as measured by equipment utilization, the processor state, implementation technology, and ease of use.
The first delivery dates for the members of the series are 910 (August, 1962), 920 (September, 1962), 925 (February, 1965), 930 (June, 1964), 940 (April, 1966), 945 (~1968), and 9300 (December, 1964).
The 910 and 920 were designed at the same time as a planned series of compatible computers which spanned a range of performance. The 910 has instructions which facilitate defining 920 instructions by software. For example, these include the multiply and divide step1 (see page 544) instructions in the 910 for programming the multiply and divide instruction in the 920.
The I/O facility evolved to a clean structure, with the potential for having a high degree of T and Ms data-transfer concurrency at a comparatively low cost. The IBM 7094 should be studied for a contrasting (more expensive) approach.
The instructions which help manipulate floating-point data are interesting and useful. The machine's ability to execute closed floating-point arithmetic subroutines is fairly good considering that the instructions are not hardwired.
The Programmed Operator (POP) instructions provide the ability to define an instruction set for efficient encoding. The idea appeared earlier in Atlas. However, the POP instruction calls subprograms in primary memory, instead of in fixed memory like Atlas.
A nice scheme1 is described for increasing the memory address space from 16,384 to 32,768 words. Other schemes which switch memory banks, like those in the PDP-8 (Chap. 5) and in the 65,384-word 7094 II (Chap. 41), tend to be less desirable and flexible.
The SDS 930 was used at the University of California (Berkeley) as the base machine for the design of the Berkeley Time Sharing System (Chap. 24). SDS later marketed the system as the SDS 940.
The 9300 was not a member of the original 910-930 series. There is almost symbolic language program compatibility. Several registers and extra memory transfer paths were added to form the 9300 from the 930. The power of the 9300 is only a factor of 2 times the 930 for simple instructions. However, the hardwired floating-point instructions in the 9300 increases the power over the 930 by a factor of almost 10 for arithmetic problems. It is hard to believe that the incompatible 9300 was a wise choice. (We suggest a more reasonable alternative could have been a two-processor 930'. The 930' processor would be a 930 but with hardwired floating-point arithmetic instructions.) The 9300 has interesting twin-mode instructions for simultaneously operating on 12-bit data pairs. The 24-bit fixed-point word is sufficient for the real-time applications for which the computer was designed.
A flaw in the series is the sharing of K's among peripheral T's and Ms's. This problem can be seen by looking at the PMS structure (Chap. 42, Fig. 2, page 546). The connection to the peripheral K from K('Channel) requires a continuous connection during the data-transfer dialogue to Mp. This structure is especially bad in the case of a slow T, for example, a typewriter. A single character transmission requires that K('W, 'Y) be assigned to the typewriter during the complete message transmission (at a connected time of 100 milliseconds/character). The problem can be avoided by placing a character memory in each slow KT. Multiple devices could then run concurrently without requiring the elaborate K('W, 'Y) to be attached to them. The structure does not preclude such an improvement.
A complete description of the input/output and interrupt system is given and should be read carefully.
1
We believe this appeared originally in the DEC PDP-1 introduced in November, 1960.542