TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

SoC Interconnect Buses: AMBA, Avalon, Wishbone, CoreConnect, OCP

The Editor's Desk
by
0
Share Facebook X LinkedIn WhatsApp Email

SoC interconnect bus: these buses are used within a chip to connect the different IP cores, such as processors, memory controllers and DMA engines, to each other and to the surrounding interface and peripheral logic. In 2005 the main choices were Altera's Atlantic and Avalon interfaces, the open WISHBONE bus from OpenCores, ARM's AMBA family (AHB, APB and ASB), IBM's CoreConnect, and the Open Core Protocol (OCP). This post lists them with links to each specification, explains what an on-chip bus does, and compares them in a deep dive.

The buses

  1. Atlantic Interface (Internet Archive copy, 2006) and the Avalon Interface Specifications, from Altera Corp. Atlantic was a standard on-chip interface for packet data, meant to let designers bridge I/O cores such as POS-PHY, UTOPIA and HyperTransport by wiring Atlantic signals together; Avalon is the interface family used by Altera's FPGA system tools.
  2. WISHBONE, from OpenCores.org. An open, public-domain interface for portable IP cores; the specification is now maintained in the FOSSi Foundation WISHBONE repository.
  3. AMBA: AHB (Advanced High-performance Bus), APB (Advanced Peripheral Bus) and ASB (Advanced System Bus), from ARM.
  4. CoreConnect Bus (Internet Archive copy of IBM's page, 2007), from IBM.
  5. Open Core Protocol (OCP), originally from the Open Core Protocol International Partnership (OCP-IP); the standard is now held by Accellera.

What an on-chip bus does

A system-on-chip is built from blocks, often bought from different vendors: a processor core, memory controllers, timers, UARTs, USB controllers and custom accelerators. Each block needs a way to read and write the others. An on-chip bus standard defines the signals, the timing of a transfer, and the rules for who may use the bus when. If every block speaks the same standard, they can be connected without custom glue logic for each pair, which is the main reason these standards exist: design reuse.

How they differ in outline

  • Shared bus versus point-to-point: early buses such as AHB and CoreConnect's PLB are shared: one master at a time drives the bus, chosen by an arbiter. Interfaces such as OCP and WISHBONE define the socket of each core and leave the interconnect (shared bus, crossbar or network-on-chip) to the integrator.
  • Fast and slow tiers: AMBA and CoreConnect both pair a high-speed bus for processors and memory with a simpler, slower bus for peripherals, joined by a bridge. This keeps slow devices from holding up fast traffic.
  • Memory-mapped versus streaming: most of these are memory-mapped (address plus data). Atlantic and Avalon streaming carry flows of data without addresses, which suits packet and DSP pipelines.
  • Licensing: WISHBONE is in the public domain; AMBA can be used without royalties; OCP is available under an Accellera license; Avalon is an open specification published by Altera.

Which one to pick

For an FPGA design, use what the vendor's tools generate, for example Avalon for Altera FPGAs. For an ASIC with an ARM processor, AMBA is the natural choice; its protocols are well documented and royalty free, and the AMBA article describes them as the de facto standard for embedded processor buses. WISHBONE suits open-source projects that want no licensing questions. OCP suited companies that wanted one core socket regardless of processor or interconnect.

Deep dive: the numbers and details

This part holds the widths, features and a bandwidth example. Skip it if the overview is all you need.

Comparison table

BusOwner and yearStructureKey facts
AMBA ASB and APBARM, 1996System bus plus peripheral busFirst AMBA version
AMBA AHBARM, AMBA 2 (1999)Shared, pipelined, arbitratedSingle clock-edge protocol; replaced ASB in most designs
AMBA AXIARM, AMBA 3 (2003); AXI4 in AMBA 4 (2010)Separate read and write channelsMultiple outstanding transactions; CHI followed in AMBA 5 (2013)
CoreConnect PLBIBMShared, arbitrated, up to 16 masters32-bit address; 32, 64 or 128-bit data; separate read and write data buses; pipelining and split transactions
CoreConnect OPBIBMBridged from PLB, multi-master32-bit address, separate 32-bit read and write buses, 8, 16 and 32-bit devices
CoreConnect DCRIBMSingle master, many slaves10-bit address, 32-bit data, no bursts; keeps control registers off the main bus
WISHBONESilicore, then OpenCores; B4 in 2010Socket; point-to-point, shared bus, data flow or crossbarPublic domain; small signal set (CYC, STB, WE, ACK plus address and data)
AvalonAlteraMemory-mapped and streaming interfacesAlso clock, reset, interrupt and conduit interfaces
OCPOCP-IP, now Accellera; OCP 3.0 is the current releaseCore socket, interconnect independentConfigurable: profiles enable only the signals a core needs

Sources: AMBA history from the AMBA article; CoreConnect details from Electronic Design, "CoreConnect: The On-Chip Bus System" (April 2, 2001); WISHBONE from the WISHBONE article and OpenCores; OCP from Accellera.

Peak bandwidth example

Peak bandwidth of a bus is data width times clock rate times transfers per clock:

bandwidth = width (bytes) x f_clk x transfers_per_clock
  • A 64-bit PLB at 133 MHz, one transfer per clock: 8 bytes x 133 MHz = 1.06 GB/s in each direction. Because PLB has separate read and write data buses, reads and writes can overlap, giving up to about 2.1 GB/s combined.
  • A 32-bit OPB at 66 MHz: 4 bytes x 66 MHz = 264 MB/s.
  • A 32-bit shared AHB at 100 MHz: 400 MB/s total, shared by all masters, since only one transfers at a time.

Real throughput is lower because of arbitration, wait states, and turnaround between reads and writes. This gap between peak and achieved bandwidth is one reason later designs moved to channel-based protocols such as AXI and to crossbars and networks-on-chip, where several masters can transfer at once.

A basic WISHBONE write

The WISHBONE classic cycle shows the handshake idea common to all these buses:

  1. The master drives the address, data and write enable (WE), and asserts CYC (a bus cycle is in progress) and STB (this transfer is valid).
  2. The slave captures the data and asserts ACK.
  3. On the clock edge where STB and ACK are both high, the transfer completes and the master can drop STB or start the next transfer.

A slave that needs more time simply delays ACK, which inserts wait states without any extra signals.

Common mistakes

  • Putting slow peripherals on the high-speed bus. Every wait state on a shared bus stalls all masters.
  • Comparing peak numbers only. Arbitration and latency matter as much as width and clock.
  • Mixing standards without checking burst and response semantics. Bridges between buses must handle differences in burst length, byte enables and error responses.

Further reading

The Editor's Desk
About the Author
The Editor's Desk
Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!