Data Communication Protocols: From Packets and Layers to SoC NoC
📡 Data Communication Protocols — From Packets and Layers to SoC NoC
A Technical Primer on Protocol Fundamentals · April 29, 2026
💡 In one line — A data communication protocol is a shared set of rules that allows heterogeneous systems to exchange data. From global internet backbones to a fingernail-sized SoC, the same core principles apply throughout: fragmentation, layering, encapsulation, and routing.
1. Why Protocols Are Necessary
A data communication protocol is a pre-agreed set of rules — a shared technical language — that allows disparate hardware and software to exchange data. Just as two engineers can collaborate across national boundaries when they share a common specification, computers, network devices, and semiconductor IP cores can only exchange meaningful bits when they follow identical conventions. Without protocols, a frame emitted by one vendor's NIC (network interface card) would be uninterpretable to every other vendor's device on the same wire.
This guide covers the subject in five steps: (1) general internet protocol concepts, (2) packets, layering, and encapsulation, (3) routing and latency optimization, (4) AMBA buses and on-chip Networks-on-Chip (NoC). Analogies and normative references are provided throughout so the end-to-end picture is clear before diving into any single topic.
2. What Is a Packet — and Why Fragment Data?
📦 2-1. Definition
A packet is the minimum transmission unit in data communications — a self-contained block carrying both a payload and the routing and control metadata needed to deliver it. The analogy: disassembling large furniture into labeled boxes for shipping, then reassembling at the destination. Each box can take a different truck route and still arrive at the right address.
⚙️ 2-2. Why Fragment?
▶ Error-recovery efficiency — A single corrupted bit in a monolithic transmission forces a complete retransmission. With packets, only the damaged unit is resent. This matters especially on high-latency or noisy links where retransmissions carry a steep time cost.
▶ Statistical multiplexing — No single sender monopolizes the link. Packets from multiple flows interleave in time, improving aggregate link utilization and enabling a shared medium to serve many concurrent users.
▶ Pipelining — Once the first packet reaches the next hop, the second packet can depart immediately. This removes the stop-and-wait penalty and reduces total transfer time proportionally to the number of intermediate nodes.
📊 2-3. Circuit Switching vs. Packet Switching
| Attribute | 🔌 Circuit Switching | 📦 Packet Switching |
|---|---|---|
| Path reservation | Dedicated end-to-end path for the session duration | Per-packet variable path; no pre-reservation |
| QoS (Quality of Service) | Deterministic; guaranteed bandwidth | Statistical; best-effort by default |
| Jitter | Negligible — fixed path, fixed timing | Present — varies with buffering and congestion |
| Key drawback | Resources wasted during silence (no data sent) | Buffer overflow and packet loss under congestion |
3. Protocol Layering — Separation of Concerns
🏗️ 3-1. Why Use Layers?
Layering is an implementation of separation of concerns: each layer exposes a well-defined interface to the layers immediately above and below it, so an implementation change within one layer does not ripple through the others. The postal analogy: changing the letter format does not affect the carrier's routing logic, and updating the truck fleet does not affect how letters are addressed. This modularity is precisely why a TCP/IP stack can run over Ethernet, Wi-Fi, cellular, or a custom on-chip fabric without changing the application layer.
📚 3-2. The OSI 7-Layer Model — Where the Standard Comes From
An important distinction: the OSI 7-layer model is an ISO standard — ISO/IEC 7498-1:1994 — not an IETF RFC. Production internet traffic uses a simplified 4-layer TCP/IP model, whose behavioral requirements are defined in RFC 1122 (Requirements for Internet Hosts, 1989). The two documents are complementary: OSI provides the conceptual reference model; RFC 1122 mandates concrete TCP/IP implementation behavior for hosts.
| Layer | Name | Primary Responsibility | PDU Name |
|---|---|---|---|
| L7 | Application | User-facing programs; raw payload origin | Data |
| L4 | Transport | Segmentation, reassembly, reliability and ordering | Segment |
| L3 | Network | Logical addressing (IP), path determination | Packet |
| L2 | Data Link | Adjacent-node MAC addressing, error detection (FCS) | Frame |
| L1 | Physical | Bit transmission via electrical or optical signals | Bit |
🔤 3-3. Key Acronyms
▶ TCP (Transmission Control Protocol) — reliable, ordered delivery via acknowledgment and retransmission. This matters because higher-layer applications (HTTP, file transfer) require byte-stream guarantees that a best-effort protocol cannot provide.
▶ IP (Internet Protocol) — logical addressing and hop-by-hop routing. Provides the "map" every router consults to forward packets toward their destination.
▶ MAC (Media Access Control) address — a 48-bit hardware identifier burned into every NIC (network interface card), used for L2 frame delivery within a single broadcast domain. MAC addresses are link-local; IP addresses are globally routable.
4. Encapsulation — Wrapping Data Per Layer
On the transmit path, each layer prepends a header containing layer-specific control fields to the data handed down from above. This process — encapsulation — produces progressively larger protocol data units (PDUs) as payload descends through the stack. The diagram below shows the canonical four-level wrapping for TCP/IP over Ethernet.
The receiver performs de-encapsulation in reverse — stripping one header per layer to recover the original payload. This abstraction is powerful: a TCP implementation never needs to know whether the underlying L2 medium is Ethernet, Wi-Fi, or a custom on-chip fabric. The layer contract isolates each concern, enabling independent evolution of hardware and software stacks.
5. Routing — How Packets Find Their Destination
Each IP packet carries source and destination IP addresses in its header. An intermediate node — a router — consults its routing table at every hop to decide which outgoing interface to use. Dynamic routing protocols (OSPF for intra-domain, BGP for inter-domain) continuously update routing tables based on link state and traffic load, so two packets with identical source and destination addresses may travel completely different paths. This path diversity allows the network to route around failures and distribute load automatically.
⚡ Latency Reduction Mechanisms
① Multiplexing — minimizes idle link time by interleaving packets from multiple concurrent flows over a shared medium.
② Pipelining — continuous packet flow with no gaps, removing the stop-and-wait latency penalty between successive transmissions.
③ Cut-through switching — a switch begins forwarding a frame as soon as it reads the destination address, before the full frame has arrived. Compared to store-and-forward, this eliminates the frame-serialization delay — most beneficial for large frames on low-congestion paths.
④ QoS / Strict Priority Scheduling — latency-sensitive traffic (VoIP, video, gaming) is dequeued ahead of bulk data. Without priority scheduling, a single large file transfer can starve a real-time flow sharing the same egress port.
⑤ TSN (Time-Sensitive Networking) — an IEEE 802.1 standard suite that provides deterministic, bounded-latency delivery for industrial automation and automotive Ethernet networks. TSN combines time synchronization (802.1AS), traffic shaping (802.1Qbv), and redundancy (802.1CB) to give Ethernet the timeliness guarantees previously requiring dedicated fieldbus systems.
6. On-Chip Data Transport: AMBA Bus vs. NoC
🔌 6-1. AMBA — The Shared Corridor Model
AMBA (Advanced Microcontroller Bus Architecture) is ARM's on-chip interconnect specification, covering AXI (Advanced eXtensible Interface), AHB (Advanced High-performance Bus), and APB (Advanced Peripheral Bus). It is fundamentally a shared medium: while one transaction holds the bus, all other IP cores wait for arbitration to complete. As the IP count scales up, bus contention grows and scheduling overhead increases — the bottleneck is structural, not fixable by tuning the bus clock alone.
🌐 6-2. NoC — A Miniature Internet Inside the Chip
A Network-on-Chip (NoC) embeds a small packet-switched network inside the SoC. Data is fragmented into flits (flow control digits) and routed over mesh, torus, or custom topologies using distributed routers. The same concepts that govern the internet — packet, header, routing table, QoS, virtual channels — apply at chip scale. This matters because a modern AI accelerator or many-core SoC may sustain dozens of simultaneous data flows; a single shared bus cannot serve them without severe head-of-line blocking.
📈 6-3. Benchmark Results (IEEE Xplore, February 2024)
※ IEEE Xplore, Feb 2024 — at high traffic and large IP count, NoC delivers approximately 2.5× throughput improvement and 30–40% latency reduction over AXI.
💬 "While AXI4 exhibits lower latency for point-to-point transfers, NoC architectures outperform in multi-core contention scenarios by distributing traffic through a mesh topology."
There is no universal winner: traffic volume and IP count are the deciding factors. The accelerating adoption of NoC in production AI accelerators and many-core SoCs reflects exactly this scaling limit of shared-bus architectures.
7. Achieving Deterministic Real-Time Delivery
Real-time does not simply mean "fast" — it means deterministic delivery: a datum must arrive within a hard deadline, every time, without exception. Packet switching is inherently probabilistic and introduces jitter (variable queuing delay). Overcoming this requires stacking complementary mechanisms across multiple layers:
The fundamental tension in modern networking is combining the efficiency of fragmented packet switching with the timeliness guarantees of dedicated circuit paths. QoS scheduling, TSN, and on-chip virtual channels are the current engineering answers to that tension — each operating at a different layer of the stack, all contributing to the same goal of bounded worst-case latency.
8. Corrections and Clarifications
▶ Correction 1 — An earlier pass treated the origin of the OSI 7-layer model ambiguously. The authoritative standard is ISO/IEC 7498-1:1994; RFC 1122 specifies behavioral requirements for TCP/IP hosts specifically. The two documents are complementary, not interchangeable — OSI defines the reference model, RFC 1122 mandates concrete TCP/IP implementation behavior.
▶ Correction 2 — An earlier pass could give the impression that NoC is universally superior to AMBA. The 2024 IEEE benchmark clarifies that AXI retains the advantage for small designs (4–8 IP cores), where NoC router overhead dominates any parallelism gain. This condition is now explicit in Section 6.
▶ Verification — No explicit contradictions were found across research passes. The two corrections above are additive clarifications; all other content is consistent.
9. Key Takeaways
✓ 1. Data is fragmented into packets for efficiency, error-recovery, and statistical multiplexing — retransmitting only the damaged unit is far cheaper than retransmitting a monolithic stream.
✓ 2. The transmitter encapsulates — prepending a header per layer — and the receiver de-encapsulates in reverse, giving each layer the illusion of direct peer-to-peer communication.
✓ 3. IP addresses, routing tables, and routers collaborate to deliver packets to their destination over dynamically chosen, failure-resilient paths.
✓ 4. Pipelining, cut-through switching, QoS scheduling, and TSN combine to enable deterministic real-time delivery over inherently probabilistic packet networks.
✓ 5. The same principles apply on-chip: AMBA (AXI/AHB/APB) for small designs; NoC for many-core, high-traffic SoCs — with approximately 2.5× throughput gain and 30–40% latency reduction at scale (IEEE Xplore, 2024).
With this framework in place, narrower topics — TCP flow control and congestion avoidance, IP subnetting and CIDR, NoC topology comparison (mesh vs. torus vs. fat-tree) — become natural extensions rather than isolated facts. The four core concepts of packet, layer, encapsulation, and routing run as a consistent thread from internet backbone routers down to the routing fabric inside a single chip.
📚 References
This content is for educational and informational purposes. For authoritative protocol behavior details, consult the primary RFC, ISO, and IEEE specifications directly.
Curating and verifying semiconductor and SoC design resources from a design and verification perspective before publishing.
This post is based on publicly available data and sources. Last updated: June 8, 2026
댓글
댓글 쓰기