Skip to content
DenseChip
Learn
PracticeCode labInterview

DenseChip

Navigate

Menu

LearnPracticeCode labInterview

Tools

All toolsSV editorCode repoResume
DenseChip

Practice, courses, and chip-design tools in one focused workspace.

LearnPracticeCode labInterviewToolsLearning paths

© 2026 DenseChip

AboutContactPrivacyTerms
Skip to lesson

Course contents

  • Architecture Metrics Reference
  • Workload Taxonomy for Architects
  • Architecture Glossary
  • Computer Architecture Cheatsheet
  • Pipeline Fundamentals14
  • Memory Hierarchy15
  • NoC and Interconnect Architecture6
    • On-Chip Networks
      • Topology Families Compared
      • NoC Latency and Bandwidth Math
      • XY Routing and Deadlock Freedom
      • Credit Flow Control and Buffers
      • Arbitration Policies and QoS
      • NoC Triage Playbook
  • Coherency and Memory Ordering9
  • Performance Analysis9
  • Accelerator Architectures6
  • SoC Architecture Tradeoffs9
  • Architecture Interview Prep14
Loading lesson

‹On-Chip Networks
  1. Topics
  2. Computer Architecture
  3. NoC and Interconnect Architecture
  4. On-Chip Networks
  5. Topology Families Compared

Computer Architecture · All levels

Topology Families Compared

Mesh, torus, ring, crossbar and hierarchical fabrics by diameter, bisection and cost.

On this page

    Five families side by side

    Diameter is the worst-case hop count between any two nodes. Bisection is the number of links crossing the worst cut, per direction.

    Average hop count assumes uniform random traffic. Real traffic has locality, so treat the average as a ceiling, not a promise.

    Family

    Diameter (N nodes)

    At N = 64

    Bisection per direction

    Cost growth

    Mesh

    2(k-1) for k by k

    14 hops

    8 links

    Wires grow with N, all short

    Torus

    k for k by k

    8 hops

    16 links

    Wires grow with N, plus long wraps

    Ring

    N/2 bidirectional

    32 hops

    2 links

    Cheapest wires, thinnest cut

    Crossbar

    1

    1 hop

    32 links

    Crosspoints grow as N squared

    Hierarchical

    Local plus bridge hops

    About 8 hops

    4 links

    Bridge bundle sets the ceiling

    Reading the worked comparison

    The block below fixes the parameters first: 64 nodes, 3 cycles per hop for router plus link, and 1 injection cycle. Every result follows by hand arithmetic.

    Diagram
    64 nodes, 3 cycles per hop (router + link), plus 1 injection cycle
    mesh 8x8:    diameter (8-1)+(8-1) = 14 hops;  zero-load 14*3+1 = 43 cycles;  bisection 8 links per direction
    torus 8x8:   diameter 4+4 = 8 hops;          zero-load 8*3+1 = 25 cycles;   bisection 16 links per direction
    ring of 64:  diameter 64/2 = 32 hops;        zero-load 32*3+1 = 97 cycles;  bisection 2 links per direction
    crossbar 64: diameter 1 hop;                 zero-load 1*3+1 = 4 cycles;    bisection 32 links per direction
    hierarchical (four 4x4 clusters, one bridge bundle of 4 links each way):
                 diameter about 6 local + 2 bridge = 8 hops; zero-load about 8*3+1 = 25 cycles;
                 bisection 4 links per direction (8 both ways)

    The torus halves mesh diameter because each dimension wraps. The wrap wires span the die edge, so timing and routing congestion pay for the hops saved.

    Cost and floorplan reality

    A 64 by 64 crossbar needs 64 times 64, or 4096, crosspoints. An 8 by 8 crossbar needs 64. Latency stays flat while area grows as N squared, which is why crossbars stop at small node counts.

    Hierarchical fabrics localize traffic only when traffic has locality. Global traffic turns every bridge into shared hotspot, so the bridge bundle must be sized from the cross-cluster matrix, not the average.

    A systolic array is also called a mesh, but it is a compute dataflow array, not a packet network. None of the rows above apply to it.

    Common mistakes

    Common pitfall

    Picking a topology from average latency alone hides the tail. Check diameter for the worst pair and bisection for the worst cut before comparing averages.

    Interview

    Likely probes: when each family wins, and what breaks the pick.

    • Mesh wins many-core with local traffic; torus buys diameter with wrap cost; ring fits modest counts; crossbar fits small fast clusters; hierarchy needs proven locality.

    • Uniform synthetic traffic never predicts burst alignment; ask for the burst model behind every average.

    Check yourself Why does the torus row show twice the bisection of the mesh row?

    Each mesh row cut crosses 8 links one way. The torus adds the wrap links in the same cut, doubling it to 16 per direction.

    Check yourself A 6 by 6 mesh reports average 3.1 hops under some workload. Uniform random theory says what, and what does the gap mean?

    Uniform random over all pairs gives about 3.9 hops, since each dimension averages (36-1)/(3*6) = 1.94 and two dimensions sum to 3.89. A measured 3.1 implies locality the reporter must state.

    Related topics

    • Latency and bandwidth math

    • Memory bandwidth and latency

    • System-level PPA tradeoffs

    Previous

    34 / 86

    Next: NoC Latency and Bandwidth Math

    Up next

    Up nextNoC Latency and Bandwidth Math

    Sign in to save progress

    Practice this lessonReport an error

    See also

    • XY Routing and Deadlock Freedom
    • Credit Flow Control and Buffers
    • Arbitration Policies and QoS