Bus
Also known as: data bus, system bus
A set of wires (or differential lanes) that carries data, addresses, or control signals between components inside a computer.
- Primary domain
- Hardware & Architecture
- Sub-category
- Printed Circuit Boards, Peripherals & Integrated Circuits
In simple terms
A bus is the highway that connects the parts of a computer. The CPU uses one bus to talk to memory, another to talk to storage, another to talk to peripherals. Each bus has a width (how many bits move at once) and a speed (how many transfers per second). When a bus is too slow, fast components sit idle waiting for data — and the bus becomes the performance bottleneck, not the processor.
The Visual Map
graph TD
CPU["CPU"] -->|"DDR5 ~100 GB/s"| RAM["DRAM"]
CPU -->|"PCIe 5.0 x16 ~64 GB/s"| GPU["GPU"]
CPU -->|"PCIe 4.0 x4 ~8 GB/s"| NVMe["NVMe SSD"]
CPU -->|"DMI / PCIe link"| PCH["Platform Controller Hub"]
PCH -->|"USB 3.2 10 Gb/s"| USBDev["USB Devices"]
PCH -->|"SATA 6 Gb/s"| SATA["SATA SSD / HDD"]
PCH -->|"I2C / SMBus"| Sensors["Sensors / EC"]
More detail
Classical computer architecture talked about three logical buses:
- Data bus — the actual values.
- Address bus — where to read/write.
- Control bus — read/write signals, clocks, interrupts.
Modern computers don’t expose these as shared multi-drop buses; they use point-to-point serial links that look the same from the outside:
- PCIe (PCI Express) connects CPUs to GPUs, NVMe SSDs, and many add-in cards. PCIe 5.0 carries ~32 GB/s per direction per 16-lane slot; PCIe 6.0 doubles this again.
- DDR / LPDDR — wide parallel links from CPU to RAM (the source of “memory bus” jargon). DDR5 dual-channel delivers ~100 GB/s.
- NVLink, Infinity Fabric — high-bandwidth links inside GPU clusters or between CCDs in a chiplet CPU. NVLink 4 carries 900 GB/s total.
- USB, Thunderbolt — external peripherals.
- I²C, SPI, CAN, SMBus — slow internal buses for sensors, microcontrollers, and automotive electronics.
The bus is often the bottleneck. A CPU may have terabytes-per-second of L1 cache bandwidth but only ~100 GB/s to main memory and a few GB/s to disk. Performance problems that look like CPU problems are frequently bus-bandwidth or bus-latency problems.
Under the Hood
On Linux, each PCI/PCIe device exposes its configuration space as a 256-byte file under /sys/bus/pci/devices/. The first 16 bytes follow the standard PCI config header:
import struct
# Simulates reading /sys/bus/pci/devices/0000:00:02.0/config (Intel UHD Graphics)
config = bytes([
0x86, 0x80, # Vendor ID: 0x8086 = Intel (little-endian)
0x9B, 0x3E, # Device ID: 0x3E9B
0x07, 0x04, # Command: Memory Space + Bus Master enabled
0x90, 0x00, # Status
0x02, # Revision ID
0x00, 0x00, 0x03, # Class code: 0x030000 = VGA controller
] + [0x00] * 244)
vid, did = struct.unpack_from('<HH', config, 0)
cls = struct.unpack_from('>I', config, 8)[0] >> 8
print(f"Vendor : 0x{vid:04X} {'(Intel)' if vid == 0x8086 else ''}")
print(f"Device : 0x{did:04X}")
print(f"Class : 0x{cls:06X} {'(VGA controller)' if cls == 0x030000 else ''}")
# Real usage (Linux only):
# with open('/sys/bus/pci/devices/0000:00:02.0/config', 'rb') as f:
# config = f.read(16)
Engineering Trade-offs
Shared bus vs. point-to-point switched:
- A shared bus (the original ISA/PCI model) is simple: one set of wires, many devices, arbitration protocol. But all devices share bandwidth, and one slow device stalls everyone.
- Point-to-point (PCIe, DDR) gives each device dedicated lanes and full bandwidth. More wires on the PCB, but no contention.
Bandwidth vs. latency:
- A wide high-bandwidth bus (DDR5: 64-bit × 2 channels) amortises overhead per transaction — good for streaming.
- High-latency buses (SATA: ~200 µs) are hidden by buffering for large sequential reads, but catastrophic for random I/O (why NVMe replaced SATA for SSDs: PCIe latency ~10 µs).
Serial vs. parallel:
- Older parallel buses (PCI, IDE) moved many bits simultaneously on many wires — fast in theory, but timing skew between lanes caused synchronisation problems at high speeds.
- Modern serial buses (PCIe, USB, SATA) run fewer wires at much higher frequencies; differential signalling cancels noise. PCIe 5.0 runs 32 GT/s per lane on a single pair of wires.
Real-world examples
- Connecting an Nvidia GPU via PCIe 4.0 x16 vs NVLink: NVLink 4.0 delivers ~900 GB/s (GPU-to-GPU) vs ~64 GB/s for PCIe — essential for multi-GPU training.
- An NVMe SSD plugged into a slower PCIe 3.0 x2 slot delivers ~3.5 GB/s; in PCIe 4.0 x4 it hits ~7 GB/s — same drive, different bus.
- USB 2.0 (480 Mb/s) vs USB 3.2 Gen 2 (10 Gb/s) for external SSDs: a 21× bandwidth difference from swapping only the bus.
Common misconceptions
- “Buses are shared wires.” Modern interconnects (PCIe, DDR, USB 3+) are almost all point-to-point with switching, not shared multi-drop buses.
- “Wider bus is always faster.” Only if traffic can use the width. Random-access workloads are typically latency-bound; a narrower, lower-latency bus often outperforms a wide, high-latency one.
Try it yourself
Compare bandwidth numbers across common bus types to see why NVMe replaced SATA and why GPU memory bandwidth matters:
python3 - <<'EOF'
buses = [
# (name, bandwidth_GB_s, latency_us)
("I2C 400 kHz", 0.00005, 10_000),
("USB 2.0 Hi-Speed", 0.060, 500),
("SATA 6 Gb/s", 0.600, 200),
("USB 3.2 Gen 2x2", 2.5, 10),
("PCIe 4.0 x4 (NVMe)", 8.0, 7),
("PCIe 5.0 x16 (GPU slot)", 64.0, 100),
("DDR5-6400 dual-channel", 102.4, 0.07),
("HBM3 (GPU on-die)", 819.0, 0.05),
]
print(f"{'Bus':<30} {'BW (GB/s)':>12} {'Latency (us)':>14}")
print("-" * 60)
for name, bw, lat in buses:
print(f"{name:<30} {bw:>12.3f} {lat:>14,.2f}")
EOF
Learn next
- DMA — Direct Memory Access lets peripherals transfer data over the bus without burdening the CPU on every byte; understanding DMA explains why bus design matters for I/O performance
- Peripheral — the devices that connect via external buses (USB, Thunderbolt, PCIe add-in cards)
- Motherboard — the PCB that physically routes the buses between CPU, RAM, and expansion slots
Read this in a learning path
All paths →This topic is part of a learning path. Start in context to keep prev/next and progress tracking.
Relationships
- Related
- Required by
Neighborhood
A visual companion to the relationships above. Click any node to visit that topic.