Intuition The One Core Idea
A chip with many cores is a tiny city, and moving data between cores is a traffic problem : the arrangement of routers and roads (which §0 will name the topology ) is the road map, and everything we measure — how far the farthest trip is, how much traffic can flow, how many roads meet at each junction — is just counting on that map. Master the picture of "routers as junctions and links as roads," and every formula in the parent note becomes obvious counting.
Before you can read a single formula in the topic note , you need the vocabulary. This page defines every symbol, word, and picture the parent leans on — starting from nothing.
Everything in a Network-on-Chip is built from exactly two things.
Definition Node, Link, and Topology
A node is a single junction on the chip — a small circuit called a router that receives data and decides where to send it next. A core, a cache, or a memory controller plugs into a node.
A link is a bundle of wires connecting two nodes directly — a one-hop road between junctions.
The topology is simply the pattern of which nodes are joined by links — the shape of the road map.
Look at the figure: black circles are nodes , black lines between them are links , and the overall pattern they form is the topology . The whole rest of the topic is just "how do we arrange these dots and lines?"
Intuition Why "junction" is the right mental image
A junction does one job: a thing arrives, and it must leave by one of a few roads. Every property we care about (latency, bandwidth) is a statement about how a packet hops from junction to junction . Hold this picture — we now put numbers on it.
N and M — how many nodes
N is a plain count of nodes in the network (or the number of nodes along one side of a grid). When a shape has two directions — like a rectangular grid — we use N for one side and M for the other. So a "4 × 4 grid" means N = 4 rows and M = 4 columns, giving N × M = 16 nodes.
Picture: count the dots. That's it. N and M are never mysterious — they are how many circles you can see.
Why the topic needs it: every formula (diameter, bandwidth) is written in terms of N , because we want to know how the network behaves as we add more cores . A formula in N tells you the future — what happens at 64 or 256 cores.
A hop is one journey across a single link — moving from one node to the neighbouring node. If a packet travels router → router → router, that is 2 hops .
In the figure, the red path goes from the top-left node to the bottom-right node in 4 hops. Each red segment is exactly one hop. Distance on a chip is not measured in millimetres here — it is measured in hops , because each hop costs a router's worth of delay.
Mnemonic Hop = one handshake
Every time a packet crosses a link it "shakes hands" with a new router. Count the handshakes = count the hops.
Now we can define the first real metric.
D
The diameter D is the largest number of hops between any two nodes, using the shortest path between them. It is the network's "worst-case commute."
Intuition Why we care about the
worst case, not the average
If you promise a customer a delivery time, you quote the slowest possible trip, not the luckiest one. Diameter tells a chip designer the maximum latency any two cores could suffer. A small D means "even the farthest cores are close."
Picture: stand at the most awkward corner, walk the shortest route to the most awkward opposite corner, and count hops. That count is D .
Let us make this concrete for a grid, so you can see exactly where the parent's mesh formula comes from.
D = ( N − 1 ) + ( M − 1 )
What we do: count the hops of the single worst trip — one far corner to the opposite far corner.
Why: in a grid there are no diagonal roads, so to move from column 1 to column N you must take one hop per column boundary, which is N − 1 hops. Likewise moving from row 1 to row M costs M − 1 hops.
What it looks like: trace the red staircase path in the figure of §2 — every rightward step and every upward step is one hop. Adding the two directions gives D = ( N − 1 ) + ( M − 1 ) .
For a 4 × 4 mesh: D = ( 4 − 1 ) + ( 4 − 1 ) = 6 hops.
In the figure the red node in the middle has degree 4 (roads North, South, East, West). A node on the edge has fewer roads, so lower degree.
Intuition Why degree = cost
Each road into a junction needs its own set of wires and its own port on the router (buffers, arbitration logic). More degree = a bigger, hotter, more expensive router. So degree is a price tag , and topologies that keep degree small and constant (like a ring with degree 2) are cheap to build. This is why the parent warns "more links ≠ always better" — you pay for every port.
This is the metric people get wrong most often, so we build it slowly.
Definition Bisection and bisection bandwidth
To bisect the network is to draw a line that splits the nodes into two halves of equal size (if the total number of nodes N is odd, exact halves are impossible, so we split as evenly as we can — one side gets ⌈ N /2 ⌉ nodes, the other ⌊ N /2 ⌋ ).
For any one such balanced cut, add up the data rate of every link it crosses. The bisection bandwidth B b is the minimum of that total over ALL possible balanced cuts — the width of the network's narrowest possible bridge between two equal halves.
The red dashed line in the figure cuts the grid down the middle. Only the links it crosses (there are 4 of them) can carry traffic between the left half and the right half. If both halves want to talk to each other at full blast, all that traffic must squeeze through those few links.
minimum over all cuts?
A network is only as strong as its weakest choke-point. If some way of splitting the nodes in half is served by very few links, an adversarial traffic pattern will target exactly that split and jam it. So we report the worst (smallest) balanced cut , not a lucky wide one. A ring, cut anywhere, only ever severs 2 links — so no matter how many nodes you add, half the chip talks to the other half through just 2 roads. That is why the parent calls the ring's bisection "terrible." A big B b means the network won't jam under heavy all-to-all traffic — see Latency and Throughput Trade-offs .
Every B b formula is written as "(number of links crossing) × B link ." So what is B link ?
Definition Link bandwidth
B link
The link bandwidth is how much data one link moves per second, in bits per second (b/s) .
Intuition Why multiply frequency by width?
Picture a link as a conveyor belt. w is how many boxes fit side-by-side on the belt; f is how many times per second the belt lurches forward one step . Boxes-per-second = boxes-per-step × steps-per-second = w ⋅ f . That is the entire derivation.
Worked example One link's rate
w = 64 bits, f = 2 GHz = 2 × 1 0 9 firings/second.
B link = 64 ⋅ 2 × 1 0 9 = 1.28 × 1 0 11 b/s = 128 Gb/s .
This is the "64 bits @ 2 GHz = 128 Gb/s" you see repeated throughout the parent note.
The formulas sprinkle a few pieces of maths notation. Here is each, from zero.
⌊ x ⌋
⌊ x ⌋ means "round x down to the nearest whole number ." So ⌊ 4 ⌋ = 4 , ⌊ 4.9 ⌋ = 4 , ⌊ 3.5 ⌋ = 3 . (Its partner ⌈ x ⌉ , used in §5, rounds up : ⌈ 3.5 ⌉ = 4 .)
Why the topic needs it: hops are whole numbers — you cannot make half a hop. The ring's diameter ⌊ N /2 ⌋ says "go at most halfway around, rounded down." For N = 8 , ⌊ 8/2 ⌋ = 4 hops.
log 2 N — "how many doublings to reach N "
log 2 N answers: starting at 1, how many times must I double to reach N ? Since 1 → 2 → 4 → 8 → 16 is 4 doublings, log 2 16 = 4 .
Intuition Why logarithms appear in the best topologies
A hypercube's diameter is log 2 N . That is astonishingly small: double the number of cores, and the worst-case trip grows by only one hop. Whenever a structure lets you halve the remaining distance each step , the number of steps is a logarithm. This is why log = "great scalability" in the comparison table.
Definition Bit, binary label, and bit-indexing
A bit is a single 0 or 1. A node's number written in bits is its binary label — e.g. node 5 is 0101. We index bits from the right, starting at 0 : the rightmost bit is bit 0 (the least-significant bit , worth 2 0 = 1 ), the next is bit 1 (worth 2 1 = 2 ), then bit 2 (worth 4 ), and so on. So in 0101 the bits that are 1 are bit 0 and bit 2.
⊕
The XOR operation ⊕ compares two bits and gives 1 if they differ, 0 if they match . Applied to two binary labels, it lights up exactly the positions where they disagree.
5 ⊕ 2 1
5 = 0101. 2 1 = 0010 (only bit 1 set). XOR flips exactly bit 1 (counting from the right, index 0): 0101 → 0111 = 7 . So node 5 connects to node 7 by "flipping bit 1."
Definition Hamming distance
The Hamming distance between two binary labels is simply the count of bit positions in which they differ — equivalently, the number of 1s in their XOR.
Intuition Why XOR (and Hamming distance) describe hypercube neighbours
In a hypercube, two nodes are neighbours exactly when their labels differ in one bit — Hamming distance 1. To travel between any two nodes you must fix every disagreeing bit, and each fix is one hop, so the number of hops equals their Hamming distance. XOR is the tool that reveals those differing bits instantly — that's why the parent uses it. See Routing Algorithms for how routers exploit this.
O ( something ) — the growth shape
O ( f ( N )) is shorthand for "as N grows large, this quantity grows roughly like f ( N ) ," ignoring constant factors. O ( N ) grows like a square root; O ( log N ) grows very slowly; O ( N ) grows in a straight line.
Why the topic needs it: the comparison table uses O ( ⋅ ) so you can compare topologies at a glance without picking a specific N . It answers "which topology wins as chips get huge?" — the theme of System-on-Chip (SoC) Design scaling.
Link = road between nodes
Diameter D = worst-case hops
Link bandwidth Blink = f times w
NoC Topologies parent note
Read it top-to-bottom: the two atoms (node, link) feed every metric, the metrics plus the maths notation feed the formulas, and the formulas are the parent topic. Power and coherence sit downstream — see Power Management in SoCs and Cache Coherence Protocols .
Cover the answers; you are ready when you can say each aloud.
What is a node, in one phrase? A router — a junction where data arrives and is sent onward.
What is a link? A bundle of wires forming a one-hop road between two nodes.
What is a topology? The pattern of which nodes are joined by links — the shape of the road map.
What is one hop? A single journey across one link, to a neighbouring node.
Define diameter D . The largest shortest-path hop count between any two nodes (worst-case commute).
Why is a mesh's diameter ( N − 1 ) + ( M − 1 ) ? No diagonals, so you pay N − 1 hops across and M − 1 hops down for the farthest corner-to-corner trip.
Define the degree of a node. The number of links touching it = number of one-hop neighbours = router cost.
What does bisection bandwidth measure? The MINIMUM, over all balanced cuts, of the total data rate of links crossing that cut.
How do you split an odd number of nodes into "halves"? As evenly as possible — one side gets ⌈ N /2 ⌉ nodes, the other ⌊ N /2 ⌋ .
Give B link and explain each factor. B link = f ⋅ w ; f = firings per second, w = bits per firing.
What does ⌊ x ⌋ do? Rounds x down to the nearest whole number.
What does log 2 N count? How many doublings from 1 reach N (why hypercube diameter is tiny).
Which bit is "bit 0"? The rightmost (least-significant) bit, worth 2 0 = 1 .
What does a ⊕ b give for two bits? 1 if the bits differ, 0 if they match — reveals hypercube neighbours.
Define Hamming distance. The count of bit positions where two labels differ = number of 1s in their XOR = hop count in a hypercube.
What does O ( log N ) mean versus O ( N ) ? Grows very slowly (great scaling) versus grows in a straight line.