Help !

I can try to remember and add something to describe those two unusual and limited production IBM products. 2938 was a follow up of the 2937, basically an improved performance unit due to more modern technology implemented. There has been also a 3838, the last one of the series: no info on that , sorry. Data-flows for 2938 and 2937 were quite similar. Principle was exactly the same.

As an IBM CE I trained on 2937 and 2938, but I mainly worked on 2938. Those units were especially designed to make calculations on arrays of data. Reason why they were almost exclusively deployed by oil research companies was because of their need to quickly process huge amount of data, in form of arrays.

Apart from oil companies, one machine had been installed at La Sorbonne University in Paris (research on pure mathematics) and another at Boeing for simulations.

Prospecting for oil was (still is?) done by working on bounced back waves following an explosion. The waves were digitally captured on tape by a number of 'geophones' (21 of them, thus a special tape drive, not IBM, with 21 tracks on an 1 inch tape with just one huge big record from beginning to the end of the recording) placed at pre-defined locations around the center of the explosion. A similar tape unit had to be present at the computer room, to read those tapes coming from the field, land or sea. The principle is very similar to sonar's installed into submarines and war ships.

The main calculation needed to process those wave shapes is FFT (Fast Fourier Transform). FFT was among the native OP codes of 2937/8, with others like Vector Multiply, Sum of Squares, Vector by Vector and Element by Element Sum, Differential Equation, and so on. All those instructions were natively built-in into the Array Processor. It was just enough to tell it what to do (OP code), where the arrays (in and out) were located into the CPU storage, how long they were and all the rest was carried out by the Array Processor, freeing the S/44 CPU to do other stuff.

Differently from any other CPU, the ALU (Arithmetic Logical Unit) of an Array Processor was performing, with a 'pipeline' layout, one basic operation on three floating point numbers: normalize if due, multiply two numbers, add the result to the third number and normalize the result. While a 'normal' ALU just adds two numbers together. The pipeline, in a given moment, could contain several series of those tern of numbers (two to be multiplied and one to be added) being processed. Sequence was: first normalize, then multiply, then add, then post-normalize. So, 4 stages, with data going on to the next stage when done and another value coming in from the previous stage. I guess the machine cycle time was 10 nanoseconds. It means that, once fully loaded, every 10 nanoseconds a normalized result of the multiply + add operation was popping out at the end of the pipeline. One billion normalize, multiply, add, normalize long precision floating point operations in 10 seconds. Not bad, for those years !!! We were at the limits of the technology, though: electric signals through copper travel at about 70% the speed of light and so signals travel about 30 cm in one nanosecond. So the length of the cables from S/44 CPU to 2937 was limited to some few meters, the two units had to be installed very close each other.

There was a micro-program to manage the process, but due to the power of the hardware, each operation was carried out in very few microcode steps.

One funny note: on the 2938's console panel , there were so many lights that the 'lamp test' switch was divided in two switches. One for upper part and one for lower part and mutually exclusive. Reason being that all lights ON at the same time would have caused an unbearable load for the lamp Power Supply! But ... to watch the console when the machine was running was very impressive .... like in the sci-fi movies!!!