Credit-Based Flow Control

In AXI, we use VALID/READY. If READY is low, the Master must wait. In a huge network (NoC), this is inefficient because the "Wait" signal takes too long to travel back to the Master.

CHI Solution: The receiver tells the sender: "I have space for 5 packets." The sender takes these 5 credits. Every time it sends a packet, it subtracts 1. When it reaches 0, it stops. When the receiver clears space, it sends more credits back.

In a large NoC, circular dependencies between message types can cause a total system freeze. CHI solves this using Independent Virtual Channels:

• REQ (Request): Master to Home Node.
• SNP (Snoop): Home Node to other Masters.
• RSP (Response): Master/Home Node completion.
• DAT (Data): Actual data movement.

What if a Home Node (HN-F) is too busy to handle a request? Instead of stalling the bus, it sends a Retry response.

• The Problem: If multiple Masters keep retrying blindly, they might starve each other (Livelock).
• The Solution: CHI uses a Protocol Credit (P-Credit) mechanism. The HN-F tracks which Masters are waiting and sends a "Credit" when space is available, guaranteeing that the Master will succeed on its next try.

When integrating CHI in a SoC, watch out for these performance killers:

• HN-F Congestion: The Home Node is the brain of CHI. If it lacks enough trackers (to track outstanding transactions), the whole system slows down.
• DVM Latency: Distributed Virtual Memory (DVM) transactions can be very slow as they must reach every node in the system. High DVM traffic can tank performance.
• Quality of Service (QoS): Without proper QoS tokens, a "noisy neighbor" (like a GPU) could hog the NoC bandwidth, starving the CPU cores.