The history of computers spans several technology generations, each defined by the dominant switching device or design paradigm: vacuum tubes → transistors → integrated circuits → microprocessors → multi-core → parallel/heterogeneous. Each transition brought dramatic improvements in speed, size, power, and accessibility.
Vacuum tube era (early 1940s)
The first electronic digital computers used vacuum tubes as switching elements. Examples: ENIAC (1946), Colossus (1943). ENIAC stored numbers in vacuum-tube flip-flops (accumulators); slightly later machines like EDSAC (1949) and UNIVAC I (1951) used mercury delay lines for working memory.
- Faster than mechanical/electromechanical predecessors.
- Large (filled rooms), power-hungry (kilowatts), unreliable (tubes burned out frequently).
- Programmed by physically rewiring or by paper tape.
This was the first generation of electronic computers. They proved that electronic computation was possible but were impractical for general use.
Transistor era (late 1940s–1950s)
The transistor, invented at Bell Labs in 1947, replaced vacuum tubes:
- Much smaller (millimeters vs centimeters).
- More reliable (no filaments to burn out).
- Lower power (milliwatts vs watts).
Second-generation computers using discrete transistors emerged in the late 1950s. Faster, smaller, and more practical electronic memory technologies (magnetic core memory) also emerged.
Integrated circuit era (1960s)
The integrated circuit (IC), developed in the late 1950s by Jack Kilby and Robert Noyce, allowed many transistors to be fabricated on a single silicon chip. Continued miniaturization gave Moore’s Law — transistor count per chip doubling every ~2 years.
ICs dramatically improved performance and reliability while reducing size and cost. Third-generation computers (1960s) were built around small-scale integrated circuits.
Microprocessor era (1970s)
The microprocessor (Intel 4004, 1971) integrated an entire CPU onto a single chip. Combined with DRAM chips for higher-capacity main memory, this enabled:
- Personal computers (Altair 1975, Apple II 1977, IBM PC 1981).
- Embedded computing in consumer products.
- Computing beyond research labs and large corporations.
Fourth-generation computers based on VLSI (Very Large-Scale Integration) reached millions of transistors per chip.
Pipelining and cache era (1980s)
Performance gains through architectural techniques:
- Pipelining: overlap fetch/decode/execute of multiple instructions to increase throughput. See Instruction execution cycle.
- Cache memory: bridge the speed gap between processor and main memory. See Cache memory.
These techniques exploited instruction-level parallelism and reduced the effective memory access time. Higher clock rates also drove performance up.
Superscalar era (1990s)
Performance increasingly relied on executing multiple instructions simultaneously within a single core:
- Superscalar execution: multiple ALUs running in parallel, dispatching multiple instructions per cycle.
- Out-of-order execution: schedule instructions based on data dependencies, not source order.
- Deeper pipelines: more stages, higher clock rates.
By the late 1990s, processors hit physical limits — clock speeds couldn’t keep climbing without unreasonable power dissipation.
Multi-core era (2000s)
When clock-speed scaling stalled, performance gains came from multiprocessing on a single chip:
- Multi-core processors: 2, 4, 8, or more cores on one chip, sharing some cache levels but with independent execution.
- Software must be multi-threaded to use the cores. Sequential programs see no speedup beyond what one core provides.
Intel and AMD shifted to multi-core in the mid-2000s. Modern mainstream desktop CPUs have 4–32 cores; HEDT/workstation parts (AMD Threadripper, Intel Xeon W) push that to 64–96 cores in a single socket, and server parts higher still.
Parallel and heterogeneous era (2010s–present)
Continued shift toward parallelism at multiple levels:
- Many-core: server CPUs with 64+ cores.
- GPUs (Graphical Processing Units): thousands of simple cores optimized for parallel work — ideal for graphics, scientific simulation, machine learning.
- Specialized accelerators: TPUs (Tensor Processing Units), FPGAs, AI chips. Different architectures for different workloads.
- Heterogeneous computing: CPU + GPU + accelerators, each handling what they’re best at.
The current trend continues toward large-scale parallelism and specialization, reflecting a shift from frequency-driven to parallelism-driven and accelerator-driven performance.
Throughline
Each generation has been driven by:
- A new switching device (tube → transistor → IC → microprocessor) or
- A new architectural technique (pipelining → superscalar → multi-core → heterogeneous).
When one approach saturates (clock speed in the early 2000s, transistor scaling now), the industry shifts to the next.
For the law that drives much of this trajectory, see Moore’s law. For the modern multi-core architecture, see Functional Units of a Computer (extended for multiple cores).