DigiFlop

Five independent channels, out-of-order completion — built for how modern SoCs actually move data.

VLSI Concept · Architecture
AMBA AXI — Advanced eXtensible Interface
Five independent channels, out-of-order completion — built for how modern SoCs actually move data.

AXI (Advanced eXtensible Interface) is the protocol every modern high-performance SoC interconnect is actually built on — the fabric connecting CPUs, caches, memory controllers, and high-bandwidth accelerators. What sets it apart from AHB isn't just speed, it's structure: instead of one shared address/data bus, AXI splits every transaction across five independent channels that all run in parallel.

The write address channel (AW) and write data channel (W) carry a write request and its payload; the write response channel (B) carries the slave's completion status back; the read address channel (AR) and read data channel (R) handle reads the same way. Because reads and writes live on entirely separate channels, a master can issue a write and a read in the same cycle without either blocking the other — something neither AHB nor APB can do, since they only have one address bus to share.

Bursts on AXI support three addressing modes: FIXED, where every beat targets the same address (useful for streaming into or out of a FIFO-mapped register); INCR, the standard incrementing burst for sequential memory access; and WRAP, which wraps back to an aligned boundary after reaching the end of the burst — again critical-word-first cache-fill behavior, just like AHB's WRAP bursts.

AXI4 extends burst length up to 256 beats for INCR bursts (versus AHB's much shorter fixed bursts) and adds byte-level write strobes (WSTRB) so a burst can write partial words cleanly. The feature that most distinguishes AXI, though, is transaction IDs: every outstanding request on the AW/AR channels carries an ID (AWID/ARID), and the corresponding response on B/R echoes that ID back (BID/RID).

This lets a master fire off many requests before any of them complete, and lets a slave — or an interconnect fanning out to multiple slaves with very different latencies — return those responses in whatever order they actually finish, as long as responses sharing the same ID stay in order relative to each other. That single mechanism is what decouples request latency from achievable throughput, which is the entire reason AXI scales to the bandwidths modern DDR/HBM memory controllers and multi-core cache-coherent interconnects need.

For the register-only end of the spectrum — the role APB used to fill exclusively — AXI4-Lite strips all of this down to single-outstanding, non-burst transactions on the same five channels, giving simple peripherals a native AXI interface without forcing the interconnect to support multiple bus standards. Because of that channel-level parallelism and the strict ordering/handshake rules for each one, AXI is also the protocol most verification engineers spend the most time on: an AXI VIP (verification IP) checking valid/ready handshakes, ID-based ordering, and protocol compliance across all five channels is a staple of SoC-level verification and a near-universal interview topic.

5 independent channelsAW (write address), W (write data), B (write response), AR (read address), R (read data) — reads and writes fully independent, all running in parallel.
HandshakeEvery channel uses a simple VALID/READY handshake — the source asserts VALID, the destination asserts READY, and the transfer happens on the cycle both are high, with either side able to stall.
Burst typesFIXED (same address every beat), INCR (sequential), WRAP (wraps to an aligned boundary — critical-word-first cache fills). Up to 256 beats per INCR burst in AXI4, with byte-level WSTRB.
Out-of-order completionAWID/ARID/BID/RID transaction IDs let a master have many requests outstanding at once, with responses returned in whatever order they finish, decoupling latency from throughput.
Why it mattersThe dominant interconnect protocol in essentially every modern SoC and FPGA design — and, because of its five-channel handshake and ordering rules, one of the most verification-heavy protocols in the AMBA family.
▶ Try the interactive version