← all systems · see it in The System
Central processing unit
The primary processor in a computer: electronic circuitry that executes the instructions of a stored program (arithmetic, logic, control and input/output), built around an arithmetic-logic unit, registers and a control unit that runs the fetch-decode-execute cycle; today one or more cores on a microprocessor chip, with caches, pipelines and other parallelism.
Properties published scienceour proposal
Sourced fields, and our own readings.
| aliases | Central processing unit (CA + CC) |
| born at | 21 June 1948 (Manchester Baby's first program); contested (C-568) |
| born precision | day |
| boundary | Inside: the control unit and its instruction decoder, the arithmetic-logic and floating-point units, the address generation and memory management units, the register file, the on-die caches, the pipeline and dispatch machinery, the clock distribution and the die-integrated power management; in a multi-core chip, each core and the shared caches. Outside: main memory and storage, input/output devices, coprocessors such as GPUs, the external clock oscillator, the power supply, the cooling, the operating system and the programs it runs, and the people who write and use them. On the line: the instruction set architecture (the contract with software, set by the designers and held in the decoder), microcode (rewritable in some designs), and the bus that carries both instructions and data. |
| nested level | Built from transistors (earlier vacuum tubes and relays) on a semiconductor, organised into logic gates, adders, latches and registers, and those into units; the CPU is one component of computer hardware and the processor of a computer. Programs, the operating system and every piece of software run on it (booked from their side as runsOn). |
| scale | technological |
Measures (26) published science
| characteristic clock: clock rate | 100 kHz to 4 MHz in early machines; tens of MHz to several GHz in microprocessors |
| characteristic clock: one instruction on the Baby | about 1.44 ms (word access 360 µs; an instruction took four times as long); about 700 instructions per second |
| Manchester Baby first program | 17 instructions, about 52 minutes, about 3.5 million operations, about 1100 instructions per second - |
| Manchester Baby build | 550 valves, 3500 W, 32 words of 32 bits - |
| memory hierarchy latencies (4.0 GHz core) | register 0.25 ns (1 cycle); L1 1 ns (4); L2 3.5 ns (14); L3 11.75 ns (47); DRAM 82.5 ns |
| uncached value wait | several hundred clock cycles |
| branch misprediction penalty | 14-25 cycles; halving mispredictions gave 13% speed |
| dual-core gain | about 50% over one core |
| per-core hand-off efficiency (derived) | about 0.75 (1.5x throughput for 2x cores) ratio |
| clock distribution share of chip power | can be over 30% |
| repeater share of gates | 50-80% of all gates in high-performance ICs |
| energy in wires and repeaters | 30-60% of all energy |
| Landauer limit at room temperature | 2.85e-21 J per erased bit, against 0.1-100 W per IC |
| forecast vs actual clock (45 nm) | 10 GHz forecast; 3-4 GHz seen |
| Alpha 21264 power | 95 W maximum; 72 W typical applications |
| cooling cost above 35-40 W | more than $1 per W per chip $/W |
| thermal throttling threshold (RPi4B) | 80 °C; cores 1.5 GHz down to 600 MHz; active cooling raised throughput up to about 90% |
| transistors per advanced microprocessor | up to 92 billion per die transistors |
| CPU in 1968 | 24 ICs of eight types, about 1000 MOSFETs each - |
| Meltdown read rate | up to 503 KB/s |
| mercurial core rate | a few per several thousand machines (Google); similar at Facebook - |
| silent corruption example | Int(1.1^53) computed as 0 on core 59 (should be non-zero); Int(1.1^52) = 142 correct |
| cosmic neutron flux (New York sea level) | about 14 neutrons/cm²/hour |
| data-transfer energy-delay saving from logic-in-memory | about 780x ratio |
| perceptron predictor on cerebellar activity | up to 85% accuracy |
| US data-centre electricity share (2011) | 2.2% |
Open contentions (7) open contention
Where sources or readings disagree. Both stay on the map.
- C-563 What causes most computer memory errors: cosmic rays or hardware faults?
booked: The map keeps both: cosmic-ray neutrons as a cause of random errors, and field data showing that hardware faults dominated in one large fleet.
the other reading: in large fleets hard errors and manufacturing defects dominate silent corruption - C-568 What was the first computer?
booked: The map names no single first. It keeps each claim (Z3, Colossus, ENIAC, Manchester Baby) with the qualifiers that make it true.
the other reading: EDSAC as the first practical stored-program electronic computer (src:wp:von-neumann-architecture); the ENIAC converted to a primitive read-only stored-program machine (16 September 1948) - C-1229 Has Moore's law ended?
booked: The map records that it has slowed since about 2010 and treats the trend as still running, keeping both 'dead' and 'continuing' claims alongside.
the other reading: Moore's law a fairly accurate predictor until 2016 (src:wp:central-processing-unit); dead (Huang 2022) / alive (Gelsinger 2022) - C-1425 Did von Neumann invent the 'von Neumann architecture'?
booked: The map uses the usual name and keeps the dispute over credit to Eckert and Mauchly.
the other reading: 'The attribution of the invention of the architecture to von Neumann is incorrect' (Eckert and Mauchly); the stored-program concept split into three paradigms (Haigh et al. 2014); Turing's universal machine and ACE; Zuse's 1936 patent applications - C-2316 Did the modern computer come from the Turing machine?
booked: The map keeps the historians' debate open.
the other reading: the stored-program concept originates in the universal machine - C-2628 How alike are brains and computer processors?
booked: The map keeps the competing readings; none is ruled.
the other reading: the brain is literally a computer under the computer-science definition, and clearly not one under the everyday definition (Richards and Lillicrap 2022); neuromorphic primitives are 'radically different' from von Neumann ones (Bartolozzi et al. 2022); shared Rentian scaling shows convergent design principles (Bassett et al. 2010) - C-2633 Can processors without a clock match clocked ones?
booked: The map keeps the competing readings; none is ruled.
the other reading: 'it is not altogether clear whether totally asynchronous designs can perform at a comparable or better level than their synchronous counterparts' (src:wp:central-processing-unit); they at least excel in simpler math operations
Partners and parts published science
Built from / builds (17)
- Computer
- Transistor
- Vacuum tube
- Semiconductor
- Clock signal
- Logic gate
- memory management unit
- Relays (historical)
- Control unit
- Arithmetic-logic unit
- Processor registers
- CPU caches (L1-L3)
- Computer hardware
- Threads competing for processor time
- Instructions competing for execution units (superscalar dispatch)
- Relays, tubes and transistors competing as the CPU's switches
- Adversarial inputs against quicksort
Sits inside / holds (4)
Trades with (valence) (15)
- Computer memory via sys:electromagnetic-field · Central processing unit gain
- Recursion: defining or computing a thing in terms of smaller versions of itself, as computer science and mathematics describe it, with its base cases, call stack, recursive data, induction, and its sightings in language and nature via sys:electromagnetic-field · Central processing unit even
- Cosmic rays via sys:cosmic-rays · Central processing unit loss
- Thermal energy (heat, temperature and heat flow) via sys:thermal-energy · Central processing unit loss
- Human (Homo sapiens) via sys:sensory-environment · Central processing unit gain
- Software via sys:electromagnetic-field · Central processing unit gain
- Operating system via sys:electromagnetic-field · Central processing unit gain
- Electricity via sys:electricity · Central processing unit gain
- Computer security via sys:electromagnetic-field · Central processing unit loss
- Compiler via sys:computer-hardware · Central processing unit even
- Semiconductor via sys:electromagnetic-field · Central processing unit gain
- Computer hardware via sys:electromagnetic-field · Central processing unit gain
- Logic gate via sys:electromagnetic-field · Central processing unit gain
- Computer-aided design via sys:electromagnetic-field · Central processing unit even
- Sorting algorithm via sys:electromagnetic-field · Central processing unit even
expressed by (17)
- Templating / Standardization
- Interoperability/Transduction (Bridging)
- Encapsulation & Modularity (Abstraction)
- Amplification/Scale-Up (Manufacture & Distribution)
- Distributed Autonomy/Decentralized Execution (Subsidiarity)
- Internal Coordination/Regulation
- Sensing/Scanning
- Uncertainty Reduction/Learning
- Risk-managed Exploration
- Trail-Marking / Externalized Memory (Stigmergy)
- Reactive Response
- Physical/Virtual Barrier
- Self-Monitoring
- Enclave/Segregation
- Triage & Prioritized Allocation
- Quality Control & Error Correction
- Bridging Supply Gaps (Store & Dormancy)
floor filled by (34)
- Computer memory
- Software
- Computer
- Cosmic rays
- Thermal energy (heat, temperature and heat flow)
- Transistor
- Operating system
- Electricity
- Compiler
- Vacuum tube
- Semiconductor
- Clock signal
- Instruction set architecture
- Logic gate
- Computer program
- memory management unit
- Relays (historical)
- Flip-flop
- Control unit
- Arithmetic-logic unit
- Processor registers
- CPU caches (L1-L3)
- Floating-point unit
- Address generation unit
- Instruction pipelining
- Branch predictor
- Speculative execution
- Out-of-order execution
- Dynamic thermal management
- Microcode
- Interrupt
- Computer-aided design
- Computer hardware
- Sorting algorithm
represents (1)
Kinds (16)
- Von Neumann architecture (one store, one path)
- Harvard architecture (rival build: separate stores for instructions and data)
- Asynchronous (clockless) computer
- Quantum computing
- Graphics processing unit
- Microprocessor
- Relay computer (Z3, Stibitz, Harvard Mark I)
- Vacuum-tube computer (ENIAC, Colossus)
- Neuromorphic hardware
- Vector architectures and GPUs
- Multi-core processor
- Microcontroller / system on chip
- Superscalar processor
- Virtual CPU (vCPU)
- Multithreading (temporal and simultaneous)
- Logic-in-memory processor
wired (3)
Notes published science
- At Facebook one CPU core answered 0 every time it was asked for the whole-number part of 1.1 to the power 53, while giving the right answer for 1.1 to the 52; nothing in the machine's logs showed a fault, and the bad answers silently deleted files (src:dixit-2021-sdc). Google sees 'a few mercurial cores per several thousand machines' (src:hochschild-2021-mercurial).
- A CPU guesses which way a program will branch, runs ahead, and throws the work away if it guessed wrong; Spectre showed the thrown-away work still leaves fingerprints in the cache, enough to read another program's secrets (src:kocher-2018-spectre).
- The same perceptron trick CPUs use to predict branches predicted real cerebellum activity up to 85% of the time (src:mittal-2018-branch-prediction), and brains and chip circuits, one evolved and one engineered, follow the same wiring law, Rent's rule (src:bassett-2010-rent).
Layers
The ways it takes in its world, where they meet, and how it acts back, each from L1 (where it meets the world) up.
the boardroom (shared floors)
- L4: The units: the control unit and instruction decoder, the ALU, FPU and AGU, the register file, the caches and the MMU, where instructions, operands, clock, power, interrupts and the core's own history meet
- L5: The microarchitecture: the instruction cycle and its parallelism (pipelining, superscalar dispatch, branch prediction, speculative and out-of-order execution, multithreading and multiple cores), with clock gating and thermal management acting on it
- L6: The instruction set architecture: opcodes, operands and addressing modes, privileged modes, interrupts and traps, the contract the hardware keeps with software
- L7: The program carried out: the stored instructions executed in sequence, with jumps for loops and conditions, their results written back to registers and memory and their I/O sent out
the central processing unit as computer engineers describe it: signals at its pins up to the program it carries out (reader's view)
- L1: The physical world at the processor: instruction and data words on the memory bus, the supply voltage, the clock signal, interrupt lines, the heat it gives off, the particles that pass through the die
- L2: Switching elements: MOSFETs today (billions per die); earlier vacuum tubes and relays
- L3: Logic and storage circuits: gates, adders, latches and flip-flops holding bits as two-valued voltages
- L4: The units: control unit and instruction decoder, ALU, FPU, AGU, register file, caches, MMU
- L5: The microarchitecture: the instruction cycle, pipelining, superscalar dispatch, branch prediction, speculative and out-of-order execution, multithreading, multiple cores; clock and thermal management
- L6: The instruction set architecture: opcodes and operands, addressing modes, privileged modes, interrupts; implemented by hardwired or microcoded decoders
- L7: The program carried out: stored instructions executed, results written to registers and memory, jumps for loops and conditions, I/O sent to devices
way in: Data (operands loaded from memory)
- L1: Data words in main memory, reached over the same bus as instructions
- L2: The MMU translates the address; the L1 data cache receives the line
- L3: Operands held in registers as binary words
way in: Electric power (supply voltage)
- L1: Supply current from the board's power delivery
- L2: The die-integrated power management module regulates voltage on demand
- L3: Rails to every switching transistor; switching draws the energy
way in: Heat (die temperature)
- L1: The die's own temperature, raised by its switching
- L2: An on-chip temperature sensor
- L3: Compared with a trigger temperature (80 °C on the Raspberry Pi 4B)
way in: Instructions (the stored program arriving from memory)
- L1: An instruction word in program memory, sent over the bus at the program counter's address
- L2: The level-1 instruction cache next to the core (on many chips; the TLB translates its address)
- L3: The program counter names the address; the fetched bits held for decoding
way in: Interrupt requests from devices and timers
- L1: A device raises a request on an IRQ line
- L2: Synchronised to the processor clock
- L3: One bit per interrupt in the mask register enables or disables it
way in: Its own past (flags, branch history, event counts)
- L1: Status bits left by previous operations (the flags register)
- L2: Branch history: outcomes of previous branches fed to the predictor
- L3: Performance counters accumulate the core's own events
way in: Microcode and firmware updates
- L1: A microcode update supplied by the vendor
- L2: Written into the rewritable memory that holds the microprogram
- L3: The microprogram turns each instruction into configuration signals over several clock pulses
way in: Other cores' writes (memory coherence)
- L1: Other CPUs writing to the shared memory
- L2: Each CPU's extra coherence hardware
- L3: Stale copies avoided, so the CPUs can share one program
way in: Stray particles (cosmic-ray neutrons, alpha particles)
- L1: Cosmic-ray neutrons (about 14 per cm² per hour at New York sea level) and alpha particles from package traces
- L2: Charge deposited in a transistor or cell
- L3: A 0 read as 1 or the reverse: a soft error
way in: The clock (time)
- L1: A periodic square wave from an external oscillator
- L2: The clock distribution network (a metal grid or H-tree) carries it to every storage element
- L3: Edges let registers and flip-flops take new values
way out: Control flow (jumps change what is fetched next)
- L3: The program counter gets the target address
- L7: A jump or conditional jump executed
way out: Heat
- L1: Heat to the cooling solution; at room temperature the Landauer floor is 2.85e-21 J per erased bit, against 0.1-100 W per chip
- L2: Dynamic power grows with clock frequency and supply voltage squared
- L5: Every switching component, clocked or not in use
way out: Input/output to devices
- L1: Signals to memory, storage, displays and network
- L4: The control unit directs data to the other devices
- L7: I/O operations of the program
way out: Microarchitectural traces (cache state, timing) left by its work
- L1: Readable by another program as access timing (a covert channel), up to 503 KB/s in Meltdown
- L4: Cache contents changed by that work stay
- L5: Wrong-path speculative work is rolled back in the nominal state
way out: Physical side effects (power draw, electromagnetic radiation, sound)
- L1: Power consumption, EM radiation and acoustic noise, used to extract keys
- L4: Computations leave observable side effects beyond their outputs
way out: Results written back (registers and memory)
- L1: To main memory over the bus
- L4: Into a register, or out through the store unit and caches
- L7: The instruction's result
way out: Throttling (slowing its own clock and voltage)
- L1: Less power drawn, less heat made
- L3: Core clock stepped down (1.5 GHz toward 600 MHz on the RPi4B)
- L5: A response chosen once the trigger fires: voltage or frequency scaling, or microarchitectural throttles
Sources (30) published science
Logic gate (Wikipedia, WME snapshot) (behind its links)
Haigh T., Priestley M., Rope C. (2014) Reconsidering the Stored-Program Concept. IEEE Annals of the History of Computing 36(1): 4-17 (authors' open copy) (behind its links)
Manchester Baby (Wikipedia, rev 1375332732, 2026-09-17, via Wikimedia Enterprise) (behind its links)
Memory hierarchy (Wikipedia, WME snapshot) (behind its links)
Kocher et al. 2018, Spectre Attacks: Exploiting Speculative Execution (arXiv:1801.01203) (behind its links)
Mittal S. A survey of techniques for dynamic branch prediction. arXiv:1804.00261 (2018) (behind its links)
Clock signal (Wikipedia, WME snapshot) (behind its links)
Markov IL (2014) Limits on fundamental limits to computation. Nature 512:147-154; arXiv:1408.3821 (behind its links)
Brooks & Martonosi, Dynamic Thermal Management for High-Performance Microprocessors, HPCA-7 (2001), doi:10.1109/HPCA.2001.903261 (behind its links)
Benoit-Cattin T, Velasco-Montero D, Fernandez-Berni J. Impact of thermal throttling on long-term visual inference in a CPU-based edge device. arXiv:2010.06291 (2020) (behind its links)
Transistor (Wikipedia) (behind its links)
Lipp et al. 2018, Meltdown, arXiv:1801.01207 (behind its links)
Hochschild PH, Turner P, Mogul JC et al. Cores that don't count. HotOS '21. doi:10.1145/3458336.3465297 (behind its links)
Dixit HD, Pendharkar S, Beadon M et al. Silent data corruptions at scale. arXiv:2102.11245 (2021) (behind its links)
Wikipedia: Soft error (behind its links)
Wikipedia: Interrupt (behind its links)
Speculative execution (Wikipedia, WME snapshot) (behind its links)
Von Neumann architecture (Wikipedia, rev 1381757639, 2026-10-24, via Wikimedia Enterprise) (behind its links)
Call stack (Wikipedia, via Wikimedia Enterprise snapshot 2026-10-01) (behind its links)
Koch S, Matveev A, Jiang Z, et al. (2019) ABC: A Big CAD Model Dataset For Geometric Deep Learning. CVPR; arXiv:1812.06216 (behind its links)
Compiler (Wikipedia, rev 1373605841, 2026-09-06, via Wikimedia Enterprise) (behind its links)
Axtmann M, Witt S, Ferizovic D, Sanders P (2020) Engineering In-place (Shared-memory) Sorting Algorithms. arXiv:2009.13569 (behind its links)
Harvard architecture (Wikipedia, rev 1345121153, 2026-03-24, via Wikimedia Enterprise) (behind its links)
Bartolozzi C, Indiveri G, Donati E. Embodied neuromorphic intelligence. Nat Commun (2022). doi:10.1038/s41467-022-28487-2 (PMC8866429) (behind its links)
Richards & Lillicrap, The Brain-Computer Metaphor Debate Is Useless: A Matter of Semantics, Front. Comput. Sci. 4:810358 (2022), doi:10.3389/fcomp.2022.810358 (behind its links)
Amdahl's law (encyclopedia) (behind its links)