Crossbar Bus Architecture: From Shared Bus to Full Matrix

Sunhong··Updated Apr 19, 2026·4 min read·192 views·Translated from Korean
Note

Chapter 4, Section 2: PL Crossbar Bus Overview

A crossbar bus is a high-performance interconnect that connects multiple masters to multiple slaves. It switches point-to-point connections dynamically, which enables multiple simultaneous transfers.

How Buses Evolved

Shared Bus

  • All devices share a single bus
  • Only one transfer at a time → bottleneck
  • Simple to implement, low bandwidth

Hierarchical Bus

  • Layered into a system bus → local bus
  • The AMBA AHB + APB structure

Crossbar (Full Matrix)

  • M masters × N slaves → M×N switch matrix
  • Every master can talk to a different slave at the same time
  • Bandwidth ↑, area and complexity ↑

Bus evolution

PL Crossbar Structure

PL (Programmable Logic) crossbar: the switch array is configured programmably.

  • Master ports: CPU, DMA, GPU, etc.
  • Slave ports: memory controller, peripherals, I/O
  • A switch at each crosspoint → arbitration + routing

PL crossbar structure

Arbitration

When several masters request the same slave, the arbiter picks one:

  • Fixed priority: priorities are fixed
  • Round-robin: fairness
  • Weighted: by bandwidth share
  • QoS (Quality of Service): real-time requirements

Arbitration

Pipelined Transaction

A bus transaction is split into an address phase and a data phase and pipelined:

  • Address and data overlap → 2x throughput
  • The basis of AMBA AHB and AXI

Pipelined transaction

Burst Transfer

Several consecutive words are transferred with a single address phase:

  • INCR4, INCR8, INCR16 (incrementing)
  • WRAP4, WRAP8 (wrap-around, for cache line fills)

Essential for cache line fills.

Burst transfer

QoS (Quality of Service)

Latency guarantees for real-time systems (video, audio):

  • Deadline-based scheduling
  • Bandwidth reservation
  • Priority queues

In AMBA AXI this is implemented with the AxQOS signals.

QoS

Scaling the Crossbar: NoC (Network-on-Chip)

With many masters and slaves, a full crossbar becomes impractical → Network-on-Chip:

  • Topologies such as mesh, torus, ring, and tree
  • Router + link structure
  • Packet-based routing

Scaling to NoC

Clock Domain Crossing

If a master and a slave sit in different clock domains, an asynchronous bridge or a FIFO is required:

  • Prevents metastability
  • Handshake protocol (req/ack)

CDC

Summary

Concept Description
Shared Bus One common bus, simple
Crossbar M×N switch matrix, multiple simultaneous transfers
Arbitration Picks a master on contention
Pipelined/Burst Techniques for higher bandwidth
QoS Real-time guarantees
NoC Interconnect for large-scale systems

The crossbar is at the heart of modern SoCs. The AMBA AXI, ACE, and CHI protocols are all built on crossbars or NoCs.

Crossbar overview