Computer Memory History

Journey and Evolution of Computer Memory | A Reflect On History

  • 1820's - 1940's | Mechanical Era

    • Punch cards & delay lines
    • Manchester Baby (1948) ran the first stored-program, but EDSAC (1949) was the first practical, regularly-usable stored-program computer — it used Mercury delay lines for memory
    • Devices: Charles Babbage’s Analytical Engine, Early IBM tabulators
  • 1950's - 1960's | Magnetic Era

    • Magnetic core: handwoven ferrite beads on grid of copper wire; each bead could be magnetized clockwise (1) or anticlockwise (0)
    • Magnetic drum: a rotating cylinder coated in magnetic material, read/write heads fixed along its tracks; a word is only accessible once the drum’s spin brings it under the head
    • Devices: Mainframes like IBM 7090 (core), IBM 650 (drum)
  • 1970's | The semiconductor revolution

    • Dynamic RAM is born (1968, Dennard/IBM) — single transistor, tiny capacitor; first commercial chip (Intel 1103) shipped 1970
    • Devices: Intel 4004 microprocessor, Apple I, Altair
  • 1970's - 1980's | ROM and Video Games Boom

    • Mask ROM – programmed during manufacture
    • PROM – programmable just one after manufacture
    • EPROM – memory could be erased by shining strong UV rays – Game cartridges were just RAM chips
    • Devices: Apple II, the IBM PC, Atari 2600 gaming console
  • 1980's | Speed vs Capacity war

    • SRAM enters the chat – its doesn’t need refresh, and is faster – But SRAM takes 4 - 6 transistors to store a single bit
    • We settle on architecture hierarchy we use today: Registers (CPU) -> SRAM (L1/L2 cache) -> DRAM (memory) -> Hard Disk (Storage)
    • Devices: Commodore 64, Macintosh, early PC’s running MS-DOS
    SRAM is actually older than DRAM

    Bipolar SRAM was invented in 1963 and MOS SRAM at 1964, both a few years before DRAM (invented 1966, patented 1968). It just stayed a niche, expensive scratchpad technology until CPUs got fast enough in the 80s to need a buffer in front of DRAM. That’s its real “arrival” — as cache, not as a chip.

  • 1990's | Synchronous Era & Rise of Modules

    • Memory & RAM used to run on their own pace
    • SDRAM (Synchronous DRAM) synced its clock speed to CPU bus
    • Memory stops being sold as loop chips, now it can be mounted into a slot on the board
    • Devices: Pentium processor & dawn of 3D gaming (Doom, Quake)
  • 2000's | Bandwidth explosion

    • DDR-RAM (Double Data Rate RAM) – instead of sending one data chunk per clock tick, DDR sends data on both rising and falling edges of the clock signal which effectively doubles speed without increasing clock frequency
    • Generations of DDR: DDR2 (lower power than original DDR, 1.8V), DDR3 (faster, 1.5V), DDR4 (dense, 1.2V); each generation changes the physical notch on the stick so you can’t put wrong one in motherboard
    • Devices: Rise of internet, Windows XP, Multicore processors
  • 2010's | The Flash Takeover

    • SSD’s using NAND flash memory killed mechanical HDD for Operating System
    • But NAND is bizarre as it stores electrons in a “floating gate” between silicon layers and to erase data you have to apply high voltage to tunnel the electrons out which degrades the silicon (meaning SSD’s have limited lifetime)
    • 3D NAND – people stopped shrinking transistors (electrons started leaking at tiny scales) and started stacking them vertically
    • Devices: Smartphones, Ultrabooks, and the Cloud
  • 2020's | The End of “memory” vs “storage”

    • HBM (High Bandwidth Memory) stacks DRAM chips directly next to processor and connects them through microscopic holes called TSVs (Through-Silicon Vias); data travels milimeters instead of inches achieving speeds of over 1 Terabyte per seconds
    • DDR5 is here and doubles burst length & improves power management by having its own builtin voltage regulator
    • Optane (RIP) & MRAM (Magnetoresistive RAM) techmologies aim to combine the speed of DRAM with non-volatility of a hard drive

The Ultimate Paradox

In the 1950’s memory was slow, expensive, and tiny, but it never forgot (core memory). Today, memory is trillions times faster, cheaper, and holds terabytes, it forgets everything the second you the power (DRAM).

We have spent 70 years trading permanence for speed. And now, in the AI era, the bottleneck is no longer CPU or storage but it is the bus between them. Engineers are now building “Processing-in-Memory” (PIM), where computation happens inside the memory chip themselves, because moving data is more expensive than calculating it.