network on chip noc
**Network-on-Chip (NoC)** is the **packet-switched communication architecture that replaces traditional shared buses or crossbar switches in complex Systems-on-Chip (SoCs), routing data packets between dozens or hundreds of distributed IP cores (CPUs, GPUs, memory controllers) using routers and scalable network topologies**.
**What Is Network-on-Chip?**
- **Definition**: A micro-network embedded directly into the silicon, functioning similarly to the Internet, but at the nanometer scale.
- **Routers**: Intelligent switching nodes placed at intersections that read packet headers and forward flits (flow control units) to the next destination.
- **Topologies**: The physical arrangement of the network (e.g., 2D Mesh, Ring, Torus, or hierarchical topologies).
- **Virtual Channels**: Multiple logical buffers sharing a single physical link, preventing routing deadlocks and prioritizing critical traffic (like memory reads).
**Why NoC Matters**
- **Scalability Limit**: Traditional shared buses (like early AMBA AHB) collapse under the extreme traffic of 10+ cores; only one device can talk at a time. NoC allows massive parallel communication.
- **Wire Delay**: In deep submicron nodes, signals cannot cross a large chip in a single clock cycle. NoC uses pipelined links, breaking the journey into multi-cycle manageable lengths.
- **Modularity**: New IP blocks can be easily attached to the NoC without redesigning global wire routing, massively accelerating SoC design cycles.
**Design Tradeoffs**
| Topology | Hardware Cost | Latency | Scalability |
|--------|---------|---------|-------------|
| **Crossbar** | Extremely High ($N^2$ wires) | Lowest (1 hop) | Very Poor (Limits at ~8-16 agents) |
| **Ring** | Low (Daisy-chained) | High (Worst-case) | Moderate (Intel CPUs use multi-rings) |
| **2D Mesh** | Moderate (Grid of routers) | Moderate | Excellent (Standard for AI accelerators) |
NoC is **the fundamental circulatory system of the many-core era** — without decentralized packet routing, scaling modern processors past a few cores would immediately choke on their own internal traffic jams.