IBM RAMAC

System versus device. RAMAC expanded to Random Access Method of Accounting and Control. The 305 was the accounting system; the 350 was its disk file. Direct record access reduced the need to sort transactions into the order of a sequential master file. This was a change in data processing, not the invention of every form of random-access memory: magnetic drums already supported addressed access. [3]

Physical organization. Noyes and Dickinson describe fifty 24-inch disks, two recording surfaces per disk, 100 tracks per surface and 500 characters per track: 50 x 2 x 100 x 500 = 5,000,000 characters. Two heads straddle a selected disk. Their carriage moves vertically; the arms move radially and retract before changing disk level. Compressed-air jets maintain head spacing. [1]

Capacity is not a byte interface. The manual separates six character-data bits from an odd-parity check. The design paper counts eight recorded positions including a space position. These boundaries describe coding, checking and recording, respectively; calling all of them usable bytes loses the distinction. [2][1]

Addressing belongs to the application. The manual organizes storage into 50,000 sectors of 100 characters. Its examples either derive an address directly from an account number or calculate a storage address and handle collisions. A two-sector employee record consumes 200 characters: sector count is not automatically application-record count. [2]

Two meanings of track. In its programmed-addressing example, the manual groups ten sectors into each of 5,000 track-address groups. The engineering paper counts 10,000 tracks on individual recording surfaces. Both descriptions yield five million characters; do not combine one convention's track count with the other's characters per track. [2][1]

The cost of an index. In the manual's illustrated layout, one of a group's ten sectors is an index to the other nine. Overflow needs another seek. The manual therefore recommends loading frequently used records first, leaving overflow space for less active records. This is an application layout choice, not a reduction in the device's raw character capacity. [2]

Checking a write. The machine compares disk readback character by character with the sending record reread into core memory. A disagreement stops execution and raises the file-error indicator. Parity checks separately test whether transferred characters have the required odd number of set bits. Neither mechanism, by itself, establishes long-term retention or recovery after an interrupted operation. [2]

Timing needs its source. The 1957 paper gives about 0.5 seconds average access and 0.75 seconds maximum; ASME's 1984 retrospective lists 0.6 seconds average. These are source-specific descriptions, not samples that can be pooled into a measured latency distribution. [1][3]