# Communication is the realization medium

Canonical: https://semiaifoundry.com/learn/atlas/crg-communication-graph/
Revision: 2026-10-07
Rights: https://semiaifoundry.com/ai-atlas/licenses.html

· Atlas v1.3.0

# Communication is the realization medium

Maps computational state, data movement and control into a graph of physical interfaces.

[Explore in Atlas](https://semiaifoundry.com/ai-atlas/#crg-communication-graph)Signal, wire, link, fabric and workload graph

Learning objectiveDraw a communication edge with a payload, receiver, timing boundary and error condition.

## The mechanism

Every physical implementation communicates: gates exchange signals, memories exchange addresses and data, dies exchange packets, and systems exchange work and checkpoints. The computation-realization graph treats these transfers as first-class objects with source, destination, encoding, rate, latency, energy and error conditions. This makes data movement visible beside arithmetic. It also explains why communication theory is useful to chip design: a representation is only operational when a receiver can distinguish it within the available noise, timing and power margins. The graph can be refined from a conceptual edge to a routed wire or protocol transaction without changing its teaching role.

1.  Represent a computational dependency as a producer-to-consumer edge.
2.  Assign an encoding and distinguishability condition to the edge.
3.  Add rate, latency, energy and error budgets at the chosen boundary.
4.  Map logical edges to physical paths, buffers, interfaces and protocols.
5.  Measure delivered useful information rather than nominal link capacity.

## Science and assumptions

### A communication edge

$$
t_{\mathrm{edge}}=\alpha+\frac{M}{\beta},\qquad E_{\mathrm{edge}}=M e_{\mathrm{bit}}
$$

Edge time is startup latency plus transferred bits divided by effective bit bandwidth. Edge energy is transferred bits times effective joules per bit.

M is bits in both relations. If a payload is given in bytes, convert it before combining it with bit bandwidth or joules per bit; the energy term omits any separately unaccounted startup energy.

Symbols and units

t\_edge

Modeled edge communication time

s

α

Fixed edge startup latency

s

M

Bits transferred at the defined edge boundary

bits

β

Effective bit-transfer bandwidth at that boundary

bits/s

E\_edge

Modeled energy associated with those transferred bits

J

e\_bit

Effective energy per transferred bit under the stated convention

J/bit

A simple edge model separates startup latency α from payload transfer M/β and scales energy with transferred bits and an effective energy-per-bit term.

Assumptions Homogeneous effective bandwidth and energy; protocol overhead, contention and topology require additional terms.

## Engineering example

For a 1 GB collective payload over a 100 GB/s effective path with 5 μs startup, the payload term is about 10 ms; a topology or contention change can dominate the same arithmetic operation.

## Distinctions and tradeoffs

A link labeled high bandwidth does not guarantee useful workload throughput when latency, contention, serialization or receiver capacity binds.

More communication can improve locality or redundancy while increasing wires, buffers, power and verification surface.

Attention, all-reduce, KV-cache traffic and checkpointing are communication graphs whose edges can bind an AI system before arithmetic does.

## Specifications and interfaces

Edge fields

Payload, encoding, rate, latency, energy, error

Values require a declared boundary and time period.

Node fields

State, storage, transform or control

A node may be logical or physical; the model labels the abstraction.

### Interfaces

State producer Input

Encoded data, control or measurement

State receiver Output

Accepted data with timing and error conditions

## Check your understanding

### Why does the CRG include communication beside computation?

1.  Because arithmetic never consumes energy
2.  Because physical states must be transferred and distinguished between operations
3.  Because all chips use the same network topology

Answer and reasoning

Because physical states must be transferred and distinguished between operations

The receiver and the path are part of a physical implementation of a computation.

## Evidence and scope

The edge equation is a bounded teaching model. Shannon’s source addresses communication limits; it does not by itself specify a chip architecture.

Use the manufacturing investigation to trace compatible inputs, assembly and operating constraints; no current standalone calculator models this lesson’s full mechanism.

## Sources

-   [Semi AI Foundry — Communication Theory of Computation Realization](https://semiaifoundry.com/updates/communication-theory-of-computation-realization/)
-   [Claude Shannon — A Mathematical Theory of Communication](https://people.math.harvard.edu/~ctm/home/text/others/shannon/entropy/entropy.pdf)

[Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/crg-communication-graph-5971d22cd051d5a1.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html)

## Continue exploring

### Foundations

-   [The computation-to-realization gap](https://semiaifoundry.com/learn/atlas/crg-problem-statement/)
-   [On-chip interconnect fabric](https://semiaifoundry.com/learn/atlas/on-chip-network/)

### Related lessons

-   [NVLink GPU interface](https://semiaifoundry.com/learn/atlas/nvlink-port/)
-   [Collective communication library](https://semiaifoundry.com/learn/atlas/nccl/)