Skip to main content
Be Bitwise

System Bus Evolution

Bus Architectures & Interconnects · Hardware Foundations track

9 min read

In this module

Why Buses Matter

Every component in a computer — CPU, memory, storage, graphics card, network adapter — needs to talk to every other one. The bus is the wiring that lets them. How fast your SSD feeds the CPU, how quickly your GPU pulls textures, how a keypress reaches the OS — all of it leans on bus architecture.

Across four decades, buses have moved from simple shared wires at a few megahertz to point-to-point serial links shoving billions of transfers per second. The evolution is worth knowing if you work anywhere near the hardware-software boundary — driver tuning, firmware analysis, attack-surface review on peripherals, all of it.

The Shared Parallel Bus Era

ISA: The Industry Standard Architecture

The ISA bus (1981) was the original PC expansion bus, born with the IBM PC. It started 8 bits wide at 4.77 MHz and grew to 16 bits at 8 MHz with the IBM AT in 1984.

ISA was a shared bus: every expansion card and the CPU sat on the same address and data lines. Only one device could drive the bus at any moment, and a simple arbitration scheme picked who got the next turn.

c
ISA Bus Characteristics:
- Data width: 8-bit (original) / 16-bit (AT)
- Clock: 4.77 MHz / 8 MHz
- Peak bandwidth: ~8 MB/s (16-bit ISA)
- Topology: Shared parallel bus
- Arbitration: Simple priority-based

ISA had no auto-configuration. IRQs and I/O addresses were set with physical jumper switches on each card. Conflicting settings — two cards trying to claim the same IRQ — drove early PC users mildly insane.

PCI: The Peripheral Component Interconnect

PCI (1992) replaced ISA with a faster, smarter bus. Auto-configuration arrived (goodbye jumpers), the data path widened, and the clock climbed.

c
PCI Bus Characteristics:
- Data width: 32-bit (standard) / 64-bit (servers, and PCI-X)
- Clock: 33 MHz (standard) / 66 MHz (PCI 2.1) / up to 133 MHz (PCI-X)
- Peak bandwidth: ~133 MB/s (32-bit/33 MHz) / ~533 MB/s (64-bit/66 MHz)
/ ~1066 MB/s (PCI-X, 64-bit/133 MHz)
- Topology: Shared parallel bus
- Arbitration: Centralised arbiter

PCI was still a shared bus. Every device hung off the same wires, and only one transaction at a time could ride them. The bus arbiter handed out access one master device per cycle.

Shared Bus Bottleneck

On a shared bus, bandwidth is divided among all connected devices. Adding more devices does not increase total bandwidth - it decreases the share available to each device. This fundamental limitation drove the transition to point-to-point topologies where each device gets a dedicated link.

Put numbers on it. Slide the device count up and watch each PCI device's share shrink, while every PCIe device keeps its whole link. Even first-generation PCIe x1 gives one device almost twice what an entire PCI bus offered. The figures are theoretical peaks, so the real PCI picture was worse.

Shared bus against point-to-point links

PCIe generation (x1 links):

PCI, 32-bit at 33 MHz: one bus for everyone

Each device
33.3 MB/s
All together
133 MB/s

PCIe 1.x: a dedicated x1 link per device

Each device
250 MB/s
All together4 links side by side, nothing shared1000 MB/s

Theoretical peaks in one direction. Real PCI loses a good share to arbitration and wait states, and PCIe loses some to packet headers. The PCIe side assumes every device has its own link to the root complex; devices behind a switch share the switch's upstream link.

The Limits of Parallel Buses

By the late 1990s, parallel buses were hitting a wall. The problems weren't logical. They were physics.

Signal Skew

A 32-bit parallel bus has 32 data wires, and the receiver expects all 32 bits to arrive simultaneously. Crank the clock and the tiny differences in wire length and capacitance between traces start mattering. Bits arrive at slightly different times. This skew caps your maximum frequency — the receiver has to wait for the slowest bit before latching anything.

Crosstalk

Adjacent parallel traces behave like coupled antennas. Crank the frequency and a signal transition on one wire induces a voltage spike on its neighbours. This crosstalk mangles data and forces engineers to either space traces further apart or back off the clock.

Pin Count

A 64-bit bus with address lines, control signals, and power needs hundreds of pins. More pins means bigger (and pricier) connectors, hairier PCB routing, and more places for solder joints to fail.

The Serial Revolution

The fix was counterintuitive: use fewer wires, not more. A serial link sends data one bit at a time (per differential pair), but at frequencies a parallel bus can't dream of.

Why Serial Wins

Differential signalling is the trick that makes this work. Instead of measuring voltage against a shared ground (single-ended), the pair encodes the signal as the voltage difference between two wires. Electromagnetic noise hits both wires the same, so the receiver just subtracts it out. That's how you get reliable signalling at multi-gigahertz rates.

c
Single-ended (ISA/PCI):
Wire: ────┐ ┌────────┐ ┌────
│ │ │ │
└──┘ └──┘
Ground: ─────────────────────────
Differential pair (PCIe/USB):
Wire+: ────┐ ┌────────┐ ┌────
│ │ │ │
└──┘ └──┘
Wire-: ────┘ └────────┘ └────
(inverse of Wire+)
Receiver reads the difference between Wire+ and Wire-

Eliminating Skew

One bit per lane per direction means no shared clock wire for the data bits to race against. Every lane is self-contained — transmit pair plus receive pair — and the receiver recovers the clock from the data stream itself. Skew doesn't vanish on a wide link: the lanes of an x16 slot still arrive at slightly different times. But the receiver measures that offset during link training and buffers each lane to line them back up (lane-to-lane deskew), which tolerates far more skew than a parallel bus's single latching edge.

Scalable Bandwidth

Want more bandwidth? Bolt on more lanes. On a shared bus, more devices means less bandwidth per device. On a serial link, more lanes means more total bandwidth — linearly.

AGP: A Dedicated Graphics Bus

By the mid-1990s, 3D graphics had gone mainstream and sharing PCI with every other peripheral became the GPU's main bottleneck. The Accelerated Graphics Port (AGP) arrived in 1997 with a dedicated point-to-point link between GPU and the chipset's northbridge.

AGP wasn't a true serial link — it was still parallel — but it was reserved for one device. The GPU got guaranteed bandwidth, no contention from network cards, sound cards, or storage controllers.

c
AGP Speeds:
AGP 1x: 266 MB/s (66 MHz, 32-bit, single data rate)
AGP 2x: 533 MB/s (double data rate)
AGP 4x: 1066 MB/s
AGP 8x: 2133 MB/s

AGP also introduced GART (Graphics Address Remapping Table): a way for the GPU to view non-contiguous system memory pages as a single contiguous buffer. That idea was a preview of the IOMMU technology a later lesson covers in depth.

AGP eventually fell to PCIe x16. More bandwidth, better scalability, and one unified interface standard instead of a graphics-only side-bus.

The Transition Timeline

The parallel-to-serial flip happened across multiple interfaces in the early 2000s:

Legacy (Parallel)Successor (Serial)Year
PCI (133 MB/s)PCIe Gen 1 (250 MB/s per lane)2003
ATA/IDE (133 MB/s)SATA (150 MB/s, then 600 MB/s)2003
Parallel SCSISAS (Serial Attached SCSI)2004
AGP (2.1 GB/s)PCIe x16 (4 GB/s Gen 1)2004

In every case the serial successor caught and overtook the parallel predecessor on bandwidth, and the gap has only widened generation by generation.

PCI Express: The Universal Successor

PCIe (PCI Express) replaced PCI, AGP, and PCI-X with one unified serial interface. It kept software compatibility with PCI — same configuration space model, same device enumeration, same driver model — while ripping out and replacing the physical and data-link layers entirely.

PCIe is packet-based, not bus-cycle-based. Data gets wrapped in packets at the transaction layer, framed at the data link layer, and serialised at the physical layer. The layered design lets each layer evolve on its own clock.

Key Concept
Packet-Based vs Cycle-Based

Legacy buses like PCI use bus cycles: the master asserts address and data lines for a fixed number of clock cycles, and the target responds. PCIe uses packets: the initiator constructs a Transaction Layer Packet (TLP) with a header and optional payload, and the link delivers it end-to-end. Packet-based protocols decouple the sender and receiver, enabling pipelining, out-of-order completion, and credit-based flow control.

Software Compatibility

PCIe gutted the physical architecture but kept the PCI software model intact. Each device still has a 256-byte (or 4096-byte extended) configuration space with vendor ID, device ID, BARs (Base Address Registers), and capability structures. The OS enumerates devices the same way it always did, and existing PCI drivers often run on PCIe devices with minimal modification.

That backwards compatibility was deliberate. Replacing the physical layer is hard enough on its own. Forcing a software-ecosystem rewrite at the same time would have been a non-starter.

Why It Matters

Bus history isn't trivia. The constraints of the early buses are why certain modern design patterns exist at all. PCI's configuration space model survives intact in PCIe. ISA's ghost is why x86 PCs still keep the 640 KB to 1 MB region of the physical address map for legacy video memory and option ROMs, and why the low end of the 64 KB I/O port space is still full of fixed legacy addresses. And the same physics that killed parallel buses — skew, crosstalk, pin count — still steers modern high-speed interconnect design today.

The next lesson dives into PCIe architecture: layered protocol stack, lane bonding, transaction types, and the topology that connects everything from NVMe drives to discrete GPUs.