Raytheon 704

Like most minicomputers of the late 1960s, the 704 was designed to be mounted in a rack mount case, with the CPU being 9U (15.75 in, 40 cm) tall.^[2][11]

The central processing unit (CPU) used 16-bit words, although it also included instructions that worked on 8-bit data, useful for working with ASCII text. Math was performed in parallel^[a] and used two's complement number format. Memory addresses were only 15-bits, allowing a maximum of 32 kilowords (64 kbytes) of directly addressable main memory.^[12]

A showcase feature was the external input/output system, handled using a separate 16-bit wide "data input/output bus", or DIO, with up to 16 devices on two physical ports.^[12] These were connected together with custom DIO cables, daisy chained up to 50 feet (15 m) in total. By using both ports on separate 50-foot chains and placing the CPU in the center, the total distance could be up to 100 feet (30 m). The end of every DIO bus had to have an electrical terminator, even on the computer ports if they were unused.^[13]

The basic unit included 4 kilowords of 18-bit memory, and a DIO controller for an ASCII terminal, normally the Teletype Model 33 or 35. This left four free card slots, and at launch, Raytheon offered a variety of plug-in expansions. These included 4 kilowords core memory modules, with enough room in the case for a total of four modules for a total of 16 kilowords.^[12]

The purpose of the 18-bit memory was to allow the use of an optional parity check module, which offloaded this task from the CPU. Two bits were used to store separate parity bits for the two bytes of data in each word. In the event an error was found, the processor was halted and a lamp illuminated on the front panel.^[14]

The hardware multiply/divide card reduced the time for a 16-bit multiply from 105 μsec to only 8, and a divide from 193 to 10.^[12]

The DIO normally had a single level of interrupt, which is fine for simple uses but not in real-time computing and similar roles where rapid hardware handling of data is a major concern. For these roles, the Priority Interrupt Expansion added an eight-level interrupt system, and a second card could be used to expand this to 16 levels.^[15] Installing this option also activated the front-panel interrupt button, which was normally inactive as the only interrupt, level 00, was assigned to the terminal.^[14]

The direct memory access (DMA) card connected up to six devices to memory, although only one device was active at a time, using a separate dedicated bus that bypassed the CPU and DIO. This was normally used for hard disk and magnetic tape support.^[14] The DMA bus also used custom cables, limited in this case to 24 feet (7.3 m) in length with only a single bus. Like DIO, the bus had to be terminated.^[16]

Finally, a Power Failsafe Option was available that was intended to stop the processor in a known-good state if it noticed the voltage from the power supply dropping, indicating an imminent power failure. An Automatic Bootstrap Module is mentioned but was not available at launch time.^[14] Sales documents from the era also mention a real-time clock and an analog-to-digital converter.^[2]

The CPU ran lock-step with the core memory, which was a typical design of the era. This limited the machine to a basic 1 μsec cycle time,^[17][b] or 1 MHz in common modern terms. The clock was a 20 MHz crystal oscillator that was divided down for the various components.^[18]

There were 72 basic instructions, plus the MPY and DIV instructions if the multiply/divide unit was installed. Instructions were generally one word long, in strong contrast to typical designs of the era which used variable length instructions depending on the addressing mode. Most instructions used the single user-visible 16-bit accumulator ACR and a 16-bit memory buffer register (MBR) used to temporarily hold operands for two-operand instructions (the other being ACR). 15-bit registers were used for the program counter (PCR) and memory address register (MAR), the later of which was used while fetching data from memory. Addresses could also be offset using the 16-bit index register IXR.^[19] Two other registers, the 5-bit EXR and 8-bit INR served special purposes.^[20]

Opcodes were generally 4-bits in length, in bits 0 through 3. The purpose of the 15-bit addresses in a 16-bit machine was to set aside one bit in the instruction format to indicate that the address was relative, normally bit 4. This left bits 5 through 15 for use as an address or constant. When used as an address, this meant it was only 11 bits long, and an additional four bits from the 5-bit EXR were added to the front to make a complete 15-bit address. This meant the memory was logically organized as a set of 2 k pages, and working with data in another page required instructions to change the EXR. All five bits of EXR were used when using byte addressing, or only four when addressing words. On top of both, the index register would be added if the index bit was turned on in the instruction.^[21]

Math instructions, ADD and SUB, always worked on 16-bit values, as did logical operations – AND, ORI ("inclusive" OR) and ORE (exclusive OR). Interesting additions were INV to invert the value and CMP to perform the two's complement. Shift and rotate instructions were available in 16 and 32-bit forms.^[22]

There were three primary comparison operators, CMW to compare words and CMB to compare bytes, and CLB to compare a byte in the accumulator with a literal byte (constant) in the address section of the instruction word. The only other support for literals was LLB which loaded a constant byte value into the accumulator.^[22] These sorts of literal (or constant) instructions are generally much more common in most processors as it avoids a memory accesses for these frequently used instructions.

Conditional branching was not supported directly, instead, there were a series of "skip" instructions that could be performed after a comparison. For instance, SAZ would skip the next instruction if the value in the accumulator was zero. To perform branches, the next instruction would normally be a jump, JMP or jump-and-store-address-in-IXR, JSX, used for subroutines. In addition to the accumulator, skip instructions were provided that read the index register, the results of a comparison and for four of the front-panel switches.^[22]

In January 1971, Raytheon announced a new add-on for the 704 system, the Array Transform Processor, or ATP. This required an entire 3U case of its own. This was essentially a vector processor dedicated to performing fast Fourier transforms (FFT), with arrays from 2 to 8192 complex numbers. A typical 2048 FFT could be accomplished in "only 150 milliseconds!" A system consisting of a 704, the ATP and enough memory to hold the data ran to about $40,000 (equivalent to $310,567 in 2024).^[23]

It was this ATP add-on which made the 704 and the later RDS500 so popular as a seismic data processing system, the ATP speeding up the many digital signal processing techniques such as filtering, deconvolution and correlation which were essential in enhancing the seismic data for interpretation.

• Raytheon 704 Technical Manual (PDF). Raytheon. July 1970.
• Dummer, G. W. A.; Thomson, F. P.; Robertson, J. Mackenzie (1971). Banking Automation: Data Processing Systems and Associated Equipment, Volume 1. Pergamon Press. ISBN 9781483160238.
• Minicomputer Review 1975. GML Corporation. 1975.