Technopedia

The technologist’s encyclopedia

Interfaces and Buses

An interface or bus is the physical and logical pathway by which information travels between two or more components of a computing system. Every transfer of data—from a register to an ALU, from a CPU to a memory chip, from a laptop to the internet—crosses at least one interface. Understanding interfaces and buses is essential whenever you are designing, debugging, or selecting hardware; choosing between DDR5 and LPDDR5, between PCIe Gen 6 and CXL, between USB4 and Thunderbolt 5, all reduce to questions about the underlying interface.

This page covers the physical layer of each interface: the medium, the signalling, the bandwidth limits, and the topology. The logical/protocol layer (framing, arbitration, error correction) is treated within each physical class below. Higher-level protocol stacks (IP, TCP, HTTP, TLS) are covered by Networks, and the information-theoretic limits that bound every medium (Shannon capacity, error correction) are covered by Communication.

Brief history

  1. 1970s – The system bus era. The IBM PC (1981) popularised a single parallel bus (ISA) shared by all peripheral cards. Every component spoke on the same wires, and arbitration was simple but bandwidth was capped.
  2. 1990s – Serialisation begins. As parallel bus widths grew, crosstalk and EMI between adjacent wires scaled with frequency, making wide parallel buses infeasible beyond ~100 MHz. Serial differential pairs became the only viable path. SPI, I²C, and UART replaced wide parallel buses for low-speed control. USB (1996) and FireWire (1995) introduced serial, point-to-point, hot-pluggable cabling to the consumer.
  3. 2000s – The on-chip explosion and the serial board. Multicore processors and System-on-Chip designs made internal interconnects (crossbars, mesh NoCs) as critical as external buses. DDR memory standards raced from DDR1 to DDR4. At the board level, the parallel-to-serial transition completed: PCIe replaced PCI, and SATA replaced PATA.
  4. 2010s–present – Coherence, chiplets, and convergence. CXL (2019) introduced cache-coherent, peer-to-peer interconnect over PCIe, a qualitative shift from shared bus to coherent fabric. The chiplet revolution (AMD EPYC 2017, Intel 2023) made die-to-die links (UCIe) a major standard, reshaping the on-chip/board-level boundary. PCIe, CXL, Thunderbolt, and USB4 share signalling layers. Optical I/O and wireless (Wi-Fi 6E, 5G) push bandwidth beyond what copper can carry at distance. DDR5 (2020) extended the memory interface race.

Categorisation

The five classes below are distinguished by the engineering regime in which the signal propagates: the scale of the conductor, the dominant loss mechanism, and the design discipline that governs it. On-chip and board-level are both electrical conduction in metal, but the physics that dominates (RC delay vs. impedance-matched transmission lines) and the tools used (transistor-level simulation vs. PCB stackup design) are fundamentally different. The “optical” row is a medium-based category that cuts across distance: optical interconnects exist at board level (silicon photonics I/O), at the cabled level (fibre between racks), and at long-haul (transoceanic cables); it is listed separately because the engineering discipline (photonics, wavelength management) is distinct from copper at every scale.

PageEngineering regimeTypical distanceTypical bandwidthTopologyRepresentative examples
On-chip interconnectsMetallic / polysilicon wires on a single die; RC-delay dominatedµm – mm (within a die)Tb/s aggregateShared bus, crossbar, ring, mesh NoCCPU bus, crossbar, mesh NoC, register file
Board-level electrical interconnectsCopper traces on a PCB and through-hole / BGA package pins; impedance-matched transmission linescm (board or package)Gb/s per lanePoint-to-point (PCIe, DDR), shared bus (I²C, SPI)DDR5, HBM3, PCIe, SPI, I²C, UART, CAN
Cabled external interconnectsShielded copper cable between systems; conductor loss + radiation dominatem – 100 mGb/s – 80 Gb/sPoint-to-pointUSB4, Thunderbolt 5, Ethernet (copper), SATA, HDMI/DisplayPort
Optical interconnectsPhotons in glass fibre or free-space optics; attenuation and dispersion dominate10 m – 1000 kmTb/s per fibrePoint-to-point, WDM multiplexedFibre Ethernet, optical I/O, data-centre interconnect
Wireless interfacesElectromagnetic waves in free space; diffraction, multipath, and path loss dominatecm – 10⁴ kmMb/s – Gb/sShared medium (CSMA/OFDMA), point-to-point (satellite)Wi-Fi, Bluetooth, NFC, Zigbee, 5G, satellite

Each class above has one explanation page per family. The cabled-external class in particular splits into five families, one page each: USB (peripheral tree), Thunderbolt (docking & expansion), Ethernet over copper (peer LAN), SATA (storage), and DisplayPort & HDMI (display).

The boundary between “board-level” and “cabled external” is not a physics boundary—both are copper transmission lines governed by the same Maxwell equations—but a system/packaging boundary: on-board traces are designed as part of a PCB stackup (controlled impedance, ground planes, via transitions), while off-board cables are designed as discrete assemblies (connector pinout, shielding, hot-plug, EMI compliance). The dominant loss mechanism also shifts: dielectric loss at the cm scale gives way to conductor loss and radiation at the m scale.

Memory interfaces (DDR5, HBM3, LPDDR5) appear in the board-level row, but the reader looking for latency and bandwidth context in the full memory hierarchy should consult Memory Hierarchy.

Die-to-die and chiplet interconnects (UCIe, AMD Infinity Fabric, Intel EMIB) operate at 1–10 mm distances through TSVs, microbumps, or organic substrate. They are subsumed under the board-level row in this table but are treated explicitly in Board-level electrical interconnects.

A single data path

A single network transfer crosses multiple classes in sequence:

CPU register → ALU → memory controller → DDR5 (board-level) → PCIe (board-level) → NIC → Ethernet cable (cabled) → fibre (optical) → router → …

Each arrow is a different interface class with its own physics, protocol, and engineering discipline. The interactive simulation below lets you trace a packet through the stack.

Details: dominant physics per class

The reason the physical medium is the right first-principles axis is that each regime is dominated by a different physical effect:

When selecting an interface for a new design, start by identifying which physical domain your signal must traverse, then consult the corresponding subpage for protocol options and trade-offs.