Protocol: AXI Interleaved vs. Out-of-Order Response
"What is the difference between interleaved and out-of-order responses?" If you are preparing for a VLSI interview, there is a good chance you will hear this question. Many people use the two terms as if they mean the same thing. They don't, and in this article we will see why.
Both concepts come from the AXI protocol, so let's first see what makes them possible.
It all starts with the ID
In AXI, every transaction carries an ID (ARID for read addresses, AWID for write addresses, and RID/BID on the way back). The ID tells the master and the slave which transactions belong together. The rule is simple:
Transactions with the same ID must stay in order.
Transactions with different IDs have no ordering requirement between them.
Everything we discuss below is just a consequence of this rule.
Out-of-Order Response
An out-of-order response means the responses come back in a different order than the requests were sent. This works at the level of a full transaction.
Say a master sends two reads: read A (ID 0) to a slow DRAM controller, followed by read B (ID 1) to a fast on-chip SRAM. If responses had to be in order, B would sit waiting behind A even though its data is ready. This is the same head-of-line blocking problem we saw in the NoC flow control article, just at the transaction level. With different IDs, the SRAM can respond first, and the master matches the data to the right request using the ID.
Here, the data of B completes fully, and then the data of A follows. Nothing is mixed.
Requests : A(ID 0) → B(ID 1)
Responses : B0 B1 B2 B3 → A0 A1 A2 A3Interleaved Response
An interleaved response goes one step further. Here, the data beats of different transactions are mixed on the same channel. This works at the level of a single beat, so you can think of it as a finer granularity.
Take the same two reads, each with 4 beats. The slave can send:
Responses : A0 B0 A1 B1 A2 B2 A3 B3Look at this carefully. Beats of A and B are alternating, but A0 → A1 → A2 → A3 are still in order, and so are B's beats. Beats of the same ID can never be reordered. This is why interleaving is useful: if A's next beat is not ready yet, the slave can use the channel for B instead of leaving it idle. If you read the virtual channel section in the flow control article, the idea should feel familiar. Many flows share one physical channel.
So, what's the difference?
Out-of-Order | Interleaved | |
Granularity | Whole transaction | Individual beat |
What changes | Order in which transactions complete | Beats of different transactions are mixed |
Needs different IDs? | Yes | Yes |
Can beats of two transactions mix? | No | Yes |
The important thing to remember: out-of-order does not need interleaving, but interleaving needs different IDs. A slave can send the whole burst of ID 1 before ID 0 without ever mixing the beats, and that is out-of-order without interleaving.
What about writes?
This is where AXI3 and AXI4 differ, and interviewers love it.
Read data interleaving is supported in both AXI3 and AXI4.
Write data interleaving was supported in AXI3 (using a WID signal on the write data channel), but it was removed in AXI4. There is no WID in AXI4, and write data must follow the same order as the write addresses.
Out-of-order write responses (the B channel) with different IDs are still allowed.
The reason for removing write interleaving was that it added complexity for slave and interconnect designers, while very few real designs got a benefit out of it.
I'll leave you guys with a question to ponder upon:
Why cannot the beats of response be out-of-order for the same transaction?Reply with the answer in the comments. See you in the next article.




Comments