Glossary

Definitions for all terms, acronyms, and variables used in this workbook.

Variables

Symbol	Full name	Definition
A	Adjustment factor	Infrastructure quality coefficient (0.5–1.0). Reflects the reliability and manageability of the WAN link. Higher A = better infrastructure.
AIW	Average AI Workload	Average Mbps consumed per effective load unit (U_eff). A composite of all concurrent AI data streams multiplied by a burst factor.
ARS	AI Readiness Score	Composite infrastructure assessment score out of 100 across five domains. Must reach 70 before AI production deployment.
B	Bandwidth	Provisioned bandwidth in Mbps on the link being evaluated.
B_gap	Bandwidth gap	B_required minus B_current. A positive value indicates an upgrade deficit.
B_min	Minimum bandwidth	Raw calculated bandwidth before safety margin.
B_rec	Recommended bandwidth	B_min multiplied by 1.5 safety margin.
B_req	Required bandwidth	Bandwidth needed to achieve IS = 3 at current load. Back-calculated from the IS formula.
CS	Cybersecurity risk factor	Dimensionless score 1–5 representing security and compliance overhead on AI traffic.
H	Human operators	Number of simultaneous human agents or operators. Multiplier: 1.0.
IS	Impact Score	Primary design validation metric. IS = (U_eff × AIW × CS × LL) / (B × A). IS ≤ 3 required for production AI.
LL	Low-latency factor	Dimensionless score 1–5 representing the latency requirement. Determines RTT budget and required inference tier.
Pod	GPU inference pod	Number of GPU inference containers or physical GPUs serving AI inference. Multiplier: 7.0.
PUO	Per-User Output	IS divided by U_eff. Per-user fairness metric. PUO > 2 indicates agent-level degradation.
RTT	Round-trip time	Total latency in milliseconds for a request to travel from user to inference and back.
U_eff	Effective load units	Weighted sum of all entity types accounting for relative network consumption. The correct input for IS and B formulas.

Technical terms

Term	Definition
ASIC	Application-Specific Integrated Circuit. The switching chip in a network switch. Cut-through ASICs (Catalyst 9K, Nexus 9K) introduce sub-microsecond latency vs legacy ASICs at 5–15 µs.
BFD	Bidirectional Forwarding Detection. Lightweight protocol for sub-second link failure detection, typically used with OSPF or BGP to accelerate failover.
CBWFQ	Class-Based Weighted Fair Queuing. QoS queuing mechanism that guarantees minimum bandwidth to each traffic class.
DPDP	Digital Personal Data Protection Act. India's data protection regulation enacted in 2023. Requires personal data of Indian citizens to be processed with specific consent and protection controls.
DSCP	Differentiated Services Code Point. 6-bit field in IP header marking traffic class for QoS treatment. EF = 46, CS5 = 40, CS4 = 32, CS3 = 24, CS1 = 8, BE = 0.
ECMP	Equal-Cost Multi-Path routing. Distributes traffic across multiple equal-cost paths. Used in campus core design for load balancing and redundancy.
FHRP	First Hop Redundancy Protocol. Generic term for HSRP, VRRP, GLBP. Provides default gateway redundancy.
FP16	16-bit floating point. Standard precision format for AI model weights. A 7B parameter model in FP16 weighs approximately 14 GB.
gNMI	gRPC Network Management Interface. Protocol for streaming telemetry from network devices. Enables sub-minute monitoring interval vs 5-minute SNMP.
GPU	Graphics Processing Unit. Repurposed for AI inference due to massively parallel architecture. NVIDIA A10G, L4, A30, A100, H100 are common inference GPUs.
gRPC	Google Remote Procedure Call. Binary protocol over HTTP/2 used by most modern AI APIs. Requires L7-aware load balancer for request routing.
HIPAA	Health Insurance Portability and Accountability Act. US regulation governing healthcare data protection. Requires specific network security controls for covered data.
HSRP	Hot Standby Router Protocol. Cisco proprietary FHRP providing default gateway redundancy with 3–10 second failover.
HTTP/2	Second major version of HTTP protocol. Uses multiplexed streams over a single TCP connection. Required for efficient AI API communication.
INT4 / INT8	4-bit and 8-bit integer quantisation formats for AI models. INT4 reduces a 7B FP16 model from 14 GB to 3.5 GB, with minor accuracy trade-off.
InfiniBand	High-speed interconnect for GPU-to-GPU communication within a cluster. 200–400 Gbps, sub-microsecond latency. Used in Tier 4 GPU clusters.
MPLS	Multi-Protocol Label Switching. WAN technology providing guaranteed bandwidth, QoS, and low jitter with provider SLA. A factor typically 0.90–0.95.
MCP	Multi-Cloud Platform. In this workbook, the four-tier AI inference architecture (Cloud, Regional Edge, Campus Edge, On-Prem) that routes workloads based on LL and CS requirements.
MIG	Multi-Instance GPU. NVIDIA technology that partitions a single GPU into independent smaller GPUs. Allows a single A30 to serve multiple workloads with isolation.
NVMe-oF	NVMe over Fabrics. Storage protocol that extends NVMe performance over a network (typically RoCEv2). Enables model weight load times of < 3 seconds vs 45 seconds on FC SAN.
NVLink	NVIDIA proprietary GPU-to-GPU interconnect. Used in Tier 4 multi-GPU clusters for distributed inference.
OSPF	Open Shortest Path First. Link-state routing protocol. With BFD, achieves 1–3 second failover on campus core links.
PCI DSS	Payment Card Industry Data Security Standard. Global standard for organisations handling card payment data. Requires network segmentation, encryption, and access controls.
PII	Personally Identifiable Information. Any data that can identify a natural person — names, Aadhaar numbers, financial accounts, health records, biometrics.
PTP	Precision Time Protocol (IEEE 1588). Hardware-assisted time synchronisation to ±1 µs accuracy. Required for AI telemetry correlation and distributed ML training log alignment.
QoS	Quality of Service. Network mechanisms for classifying, prioritising, and guaranteeing bandwidth for different traffic classes.
RAG	Retrieval-Augmented Generation. AI architecture that combines a language model with a vector database of documents. The LLM queries the database to ground its responses in factual content.
RoCEv2	RDMA over Converged Ethernet version 2. High-speed, low-latency network protocol for storage (NVMe-oF) and GPU communication.
SD-WAN	Software-Defined WAN. Overlay technology providing application-aware path steering across multiple transport links. A factor 0.82–0.90 for managed deployments.
SGT	Security Group Tag. Cisco TrustSec tag applied to packets identifying the source security group. Enables dynamic access policy without ACL-per-VLAN complexity.
SLM	Small Language Model. AI language model with 1–3 billion parameters. Suitable for Tier 4 on-prem deployment due to lower compute and storage requirements.
STT	Speech-to-Text. AI function converting audio to text in real time. Generates a continuous 0.3–0.5 Mbps stream per agent during calls. LL = 4 requirement.
TLS 1.3	Transport Layer Security version 1.3. Current encryption standard for AI API traffic. Requires NGFW with hardware TLS 1.3 decryption for DLP inspection.
VXLAN	Virtual Extensible LAN. Layer 2 overlay protocol enabling micro-segmentation across a shared IP fabric. Required for AI workload isolation in enterprise networks.
ZTNA	Zero Trust Network Access. Security model where no implicit trust is granted based on network location. AI inference pods should be ZTNA-enforced — access granted per workload, not per VLAN.

Additional terms (Chapter 9 — IS Layer-by-Layer)

Term	Definition
B1–B9	The nine bandwidth measurement points from agent access port (B1) to in-house DC fabric (B9). IS must be calculated at each point independently.
CIR	Committed Information Rate. The guaranteed minimum bandwidth on an MPLS circuit that the provider must deliver under any conditions. Use CIR as B when calculating IS for MPLS paths.
SD-WAN aggregate	When SD-WAN combines two circuits in active-active mode, the effective B is the sum of both circuit bandwidths, with a blended A factor.
WAN AIW	The portion of AIW that actually traverses the WAN link. After MCP tiering, WAN AIW = only Tier ½ cloud traffic (typically 2.0 Mbps vs full AIW of 6.8–12.0 Mbps).
A factor degradation	The reduction in A factor caused by network congestion, path quality issues, or contention. Monitoring IS in real time detects A factor degradation before it causes service failure.
Dual 25G port-channel	Two 25G links bonded via LACP into a 50G aggregate. IS in normal operation = 0.75. IS in failover (one link down) = 1.50. Both acceptable for campus AI.
Port-channel (LACP)	Link Aggregation Control Protocol. Bonds multiple physical links into a single logical interface with higher bandwidth and built-in redundancy.
Active-active SD-WAN	Both WAN circuits carry traffic simultaneously. B = sum of both circuits. Traffic steering policy (DSCP-based) determines which flows use which circuit.
Failover IS	The IS score when the primary WAN path fails and the secondary path carries the full load. Failover IS must remain below 3 for 99.9% availability design to hold.
5G ISP	A 5G-technology internet service provider link — fibre-connected infrastructure delivered to premises via 5G radio. Not the same as 5G cellular modem backup. Much higher capacity (1–10 Gbps).