AMBA CHI · Module 22 · CHI Misconceptions Engineers Have

“CHI Is Just AXI With Coherency”

The most common CHI myth: that it is AXI with a coherency layer bolted on. That description is almost exactly true — of ACE, the AXI Coherency Extensions, which are AXI's channels plus coherency signals. But CHI is not ACE. CHI exists because bolting coherency onto AXI's channel-based model did not scale, so Arm designed a new protocol: layered (Protocol, Network, Link), packet- and flit-based with credited flow control, a Request/Home/Subordinate node model, and directory coherence. None of it comes from adding a layer to AXI. The kernel of truth is real — both are AMBA, both memory-mapped, and CHI bridges to AXI — but reasoning about CHI with an AXI model is how CHI designs and debugs go wrong.

Foundation14 min readAMBA CHIMisconceptionAXIACELayered Protocol

Module 22 · Chapter 22.1 · CHI Misconceptions Engineers Have

Project thread — Module 21 closed the checklists. Module 22 refutes myths; 22.1 is the AXI conflation, 22.2 is broadcast snooping.

1. The Misconception

The belief: "CHI is just AXI with a coherency layer added on top." It shows up whenever an engineer new to CHI reaches for a familiar frame — AXI is the AMBA protocol they know, coherency is the new thing CHI adds, so CHI must be AXI + coherency. It is one of the most common and most consequential CHI misconceptions, because it plants an AXI mental model at the root of everything that follows — and that model is wrong for CHI.

2. Why It's Tempting

The belief is tempting because it is almost exactly true — of a different protocol.

ACE really is AXI + coherency. ACE (AXI Coherency Extensions) is literally AXI's channels with coherency signals and snoop channels added. The description fits ACE perfectly — and engineers conflate CHI with ACE.
Both are AMBA. AXI and CHI are both Arm AMBA interconnect standards, both memory-mapped — so they feel related.
Coherency is "the new part." From an AXI starting point, coherence looks like an addition, not a redesign — so "add a layer" feels natural.

The seed is real; the conclusion is wrong, because CHI is the protocol Arm built to replace the AXI-extension approach — not another extension of it.

3. Key Terms

4. Previous Chapter Connection

This chapter revisits the layered architecture (Chapter 4.1) and the why-CHI-was-created story (Chapter 3.8) — but as a refutation. The layers (Protocol / Network / Link), the node model (Chapter 5), the packet/flit transport (Chapter 4.6), and directory coherence (Chapter 15.3) are exactly the things the "AXI + coherency" frame erases.

Where Module 3 explained the walls AXI-style coherence hit — broadcast snoops, a shared medium — this chapter names the belief those walls disproved: that you can get coherent scale by extending AXI. You cannot; that is why CHI exists. And it is the foundation misconception, because every later Module 22 myth (snooping scales, directory means no snoops) is easier to fall for if you start from an AXI model.

5. The Claim, Precisely

State the myth falsifiably, so it can be tested.

Claim: "CHI is AXI, plus an additional coherency layer — the same channels and transport as AXI, with coherence bolted on."

This is testable. If true, CHI would share AXI's five channels, AXI's address/data transport, and AXI's handshakes, differing only by an added coherence mechanism. So the test is simple: is CHI's transport AXI's transport? It is not — CHI is packet- and flit-based, layered, and node-modelled, none of which is AXI. The claim fails its own test.

6. Why It's Wrong — The Core

CHI is not AXI with an added layer; it is a different protocol architecture.

Different transport. AXI has five independent channels (AW, W, B, AR, R) carrying address/data. CHI moves packets as flits over a layered link — not AXI channels.
Layered, not flat. CHI separates Protocol (coherence), Network (routing by node ID), and Link (flits + credits). AXI has no such layering.
Node model. CHI defines RN / HN / SN with a Home that orders and tracks sharers. AXI has masters and slaves, no home, no directory.
Directory coherence. CHI supports directory-based, targeted snooping. AXI has no coherence at all; ACE adds broadcast snoops.

You cannot reach CHI by adding anything to AXI — the transport, layering, node model, and coherence model are all replaced.

7. Engineering Diagram — CHI's own layered stack

Figure 1 — CHI is its own layered protocol stack, not a coherency layer on AXI. The Protocol layer owns coherence and transactions; the Network layer routes packets by node ID across the fabric; the Link layer moves flits with credit-based flow control; the Physical layer signals on wires. This layering — and the packet/flit transport beneath it — is what 'AXI plus coherency' erases. AXI has none of these layers; it is a flat set of address and data channels.

CHI is a layered, packet/flit-based protocol with its own transport — none of these layers exists in AXI. "AXI + coherency" flattens this stack back to AXI's channels and erases the packets, flits, credits, routing, and node model that define CHI.

8. The Mechanism

What actually distinguishes CHI's architecture from "extended AXI".

Aspect	AXI (and ACE = AXI + coherency)	CHI (ground-up)
Transport	five address/data channels	packets as flits
Layering	flat (no layers)	Protocol / Network / Link
Routing	point-to-point / interconnect	by node ID across a fabric
Flow control	valid/ready handshakes	credit-based (L-Credits)
Agents	masters / slaves	RN / HN / SN node model
Coherence	none (AXI) / broadcast (ACE)	directory-capable, targeted

The rule to carry: CHI's coherence is inseparable from its layered, packet-based, node-modelled architecture — it is not a module you could bolt onto AXI's channels. ACE tried the bolt-on approach; CHI is the redesign that replaced it.

9. The Kernel of Truth

Steelman the belief — what is genuinely right about it.

ACE genuinely is AXI + coherency. The instinct is correct for ACE — the previous-generation coherent protocol was AXI extended. The myth mislabels CHI with ACE's true description.
Shared lineage. CHI and AXI are both AMBA, both memory-mapped, and share concepts (outstanding transactions, IDs) — the family resemblance is real.
CHI interoperates with AXI. A CHI system bridges to AXI subordinates through an I/O home node (HN-I) — so AXI and CHI coexist in real SoCs.

The truth: the belief is right about the family and about ACE, and right that AXI and CHI interoperate — it is only wrong in equating CHI's architecture with extended AXI.

10. A Concrete Counterexample

Try to build "AXI + coherency" and watch what happens.

Start with AXI. Five channels, no coherence, masters and slaves.
Add coherency the obvious way. Bolt on snoop channels and coherence signals so masters can snoop each other. You have just built ACE — this is exactly what ACE is.
Scale it. ACE's snoops are broadcast over a shared interconnect — every coherent access snoops everyone. At a handful of masters it works; at dozens it collapses (the snoop storm, Chapter 3).
Hit the wall. The shared-medium, broadcast model cannot scale — the very wall Module 3 documented.
What Arm actually did. Not "extend AXI again" — they built CHI: a layered, packet-based protocol with a mesh, directory coherence, and targeted snoops. A different architecture, because the bolt-on had run out.

The counterexample is the history: "AXI + coherency" is ACE, ACE did not scale, and CHI is the ground-up answer — proof that CHI is not the bolt-on.

11. What's Actually True

The correct model, stated plainly.

CHI is a distinct, layered protocol — Protocol / Network / Link — not AXI with an added layer.
Its transport is packets and flits with credit-based flow control, not AXI channels.
It has a node model (RN / HN / SN) with a Home that orders and holds the directory.
Its coherence is directory-capable and targeted, the thing that lets it scale where ACE could not.
"AXI + coherency" describes ACE — the protocol CHI was built to succeed.

12. Why The Distinction Matters

Believing the myth has concrete costs.

You reason with the wrong model. Expecting AXI channels, AXI ordering, AXI handshakes in CHI leads to wrong mental simulations of behavior (setting up Chapters 22.4, 22.5).
You miss the layering. Debugging a link-layer credit issue with an AXI channel model fails — you look for the wrong thing (Chapter 21.2/21.5).
You mis-expect scaling. Assuming AXI/ACE broadcast behavior makes you fear a snoop storm CHI's directory avoids — or, worse, assume broadcast where CHI is targeted (Chapter 22.2).
You mis-architect bridges. Not seeing the HN-I → AXI boundary as a protocol boundary leads to integration bugs.

13. Design Implications

What to do once the myth is dropped.

Model CHI as layered — reason at the Protocol layer for coherence, Link layer for flow control, separately.
Think packets and flits, not AXI beats — transport is credited flits, not valid/ready channels.
Center the Home node — ordering and the directory live at the HN, a concept AXI has no analog for.
Treat AXI↔CHI as a bridge — an HN-I boundary between two different protocols, not a continuation of one.
Expect directory-scale behavior — targeted snoops, not broadcast (carry into Chapter 22.2).

14. The Trap

Trap: carrying an AXI mental model into CHI

The real damage of "CHI is AXI + coherency" is not the label — it is the AXI mental model it installs. Once you believe CHI is extended AXI, you reason about it as AXI: you expect five channels, so you look for AW/W/B/AR/R and miss that CHI moves packets as flits; you expect AXI ordering rules, so you mis-predict what CHI guarantees (Chapter 22.4); you expect valid/ready handshakes, so you miss credit-based flow control and mis-debug a credit stall (Chapter 21.2); you expect broadcast coherence (from ACE), so you either fear a snoop storm CHI avoids or wrongly assume broadcast where CHI is directory-targeted (Chapter 22.2). Each error is invisible as long as the design stays simple, and surfaces exactly when you need the CHI-specific mechanism — the layer, the flit, the credit, the home, the directory — that your AXI model does not contain. The label is harmless; the model it seeds quietly corrupts every CHI decision you make.

The fix: hold two models. AXI is channels, handshakes, no coherence; ACE is AXI + coherency; CHI is a layered, packet/flit, node-modelled, directory-capable protocol of its own. When reasoning about CHI, reason as CHI — layers, packets, flits, credits, homes, directory — never as AXI wearing a coherence hat.

CHI = ACE. No — ACE is AXI + coherency; CHI is the ground-up successor. This is the root of the myth.
CHI channels = AXI channels. No — CHI has packet channels (REQ/RSP/SNP/DAT) as flits, a different concept from AXI's AW/W/B/AR/R.
CHI uses valid/ready like AXI. No — CHI uses credit-based (L-Credit) flow control at the link layer.
Coherence is a bolt-on layer. No — CHI's coherence is woven through the Protocol layer and the node model, not an add-on.
AXI knowledge is useless for CHI. No — the kernel of truth: shared AMBA concepts and AXI bridging still matter; just don't equate the architectures.

16. Answering It In An Interview

If an interviewer says "isn't CHI just AXI with coherency?" — the strong answer:

Name the confusion. "That describes ACE — AXI Coherency Extensions — not CHI. CHI is a ground-up protocol."
Give the architecture. "CHI is layered — Protocol, Network, Link — and packet/flit-based with credit flow control, not AXI's channels."
Give the why. "It exists because bolting coherency onto AXI (ACE) hit a broadcast/shared-medium wall; CHI uses a mesh and directory coherence to scale."
Concede the kernel. "Both are AMBA and memory-mapped, and CHI bridges to AXI via an HN-I — so they interoperate."

That answer distinguishes CHI from ACE, shows you know why CHI exists, and concedes the true part — the signature of understanding, not recital (Chapter 21.6).

17. Key Takeaways

"CHI = AXI + coherency" describes ACE, not CHI.
CHI is a ground-up, layered (Protocol/Network/Link), packet/flit-based coherent protocol.
It has a node model (RN/HN/SN) and directory coherence — none of it an AXI add-on.
CHI exists because extending AXI (ACE) hit a broadcast/shared-medium scaling wall.
The kernel of truth: shared AMBA lineage, memory-mapped, and CHI bridges to AXI.
The trap is carrying an AXI mental model into CHI; reason as CHI.

18. Quick Revision

"CHI Is Just AXI With Coherency" — refuted. The belief is tempting because it is almost exactly true — of ACE, the AXI Coherency Extensions, which literally are AXI's channels plus coherency and snoop channels. CHI is not ACE. CHI is a ground-up, layered protocol — a Protocol layer (coherence, transactions, RN/HN/SN, snoops), a Network layer (routes packets by node ID across a fabric), and a Link layer (moves flits with credit-based flow control) — that is packet- and flit-based, has a node model (Request/Home/Subordinate, with a Home that orders and holds the directory), and supports directory-based, targeted coherence. None of that comes from adding a layer to AXI; the transport, layering, node model, and coherence model are all different. CHI exists precisely because bolting coherency onto AXI's shared, channel-based model (ACE) hit a broadcast/shared-medium wall that would not scale — so Arm built a new protocol on a mesh with directory coherence, not another AXI extension. The kernel of truth is real: both are AMBA, both memory-mapped, and CHI bridges to AXI subordinates through an HN-I — they interoperate. The trap is not the label but the AXI mental model it installs: expecting AXI channels, ordering, and handshakes, and missing the packets, flits, credits, layers, home, and directory that actually govern CHI — the source of mis-design and mis-debug. Reason as CHI, not as AXI with a coherence hat. Next, 22.2 refutes the belief that snooping scales forever.

Coming Next

Chapter 22.2 — "Snooping Scales Forever". One AXI-era assumption survives into CHI thinking: that you can always just snoop everyone. Chapter 22.2 refutes "snooping scales forever" — why broadcast snooping (the ACE model) collapses as agents multiply, why the snoop storm is a real scaling wall, and how CHI's directory turns broadcast into targeted snooping so coherence scales — the mechanism the AXI mental model from this chapter hides.