graph neural network is a neural architecture that learns from nodes, edges, attributes, and graph structure through neighborhood message passing. GNNs model molecules, circuits, social systems, fraud, knowledge graphs, traffic, and chip placement but create irregular memory and sparse-compute challenges for hardware.

Message passing. At each layer a node gathers transformed messages from neighbors, aggregates them with a permutation-invariant operation, and updates its hidden state. Stacking layers expands the receptive field; a readout produces node, edge, or graph predictions. Normalization handles degree variation. Oversmoothing can make deep node states indistinguishable, oversquashing compresses too much distant information, and heterophily breaks assumptions that connected nodes are similar.

Architecture families. GCN applies normalized graph convolution and is a strong baseline. GAT learns attention weights over neighbors. GraphSAGE samples and aggregates neighborhoods for inductive scaling. GIN uses an expressive sum-based update linked to graph-isomorphism tests. Relational and heterogeneous GNNs use type-specific transforms; equivariant networks preserve geometric symmetries for molecules and physical systems; graph Transformers introduce global or structured attention.

Systems and hardware. Real graphs have skewed degree, sparse adjacency, changing batches, and poor locality. Sampling reduces work but creates random access and data-loader overhead. Feature gathering is memory-bound, while dense transformations are compute-bound, so execution alternates regimes and underutilizes GPUs. Partitioning, caching high-degree nodes, compressed adjacency, fused gather-reduce kernels, minibatch pipelines, and distributed communication scheduling determine throughput.

Applications. Molecular GNNs predict properties and forces; fraud systems propagate account and transaction evidence; recommenders learn user-item graphs; knowledge graphs support link prediction. EDA represents netlists, timing graphs, placement neighborhoods, and routing interactions, enabling congestion, timing, or placement prediction. Models must avoid temporal leakage, respect causality and split by entities, and quantify uncertainty before affecting costly physical decisions.

Evaluation and deployment. A production implementation begins with explicit terminal conditions, operating ranges, loading, accuracy, noise, latency, efficiency, area, cost, lifetime, and fault behavior. Schematic or architectural models establish feasibility; extracted, package, board, thermal, and control-loop models then reveal interactions hidden by ideal sources and loads. Verification spans process, voltage, temperature, mismatch, aging, startup, shutdown, overload, brownout, and recovery. Teams should define measurement bandwidth, observation point, stimulus, pass limit, guard band, and statistical confidence before simulation. Layout review covers current return, thermal gradients, matching, parasitic coupling, electromigration, voltage stress, latch-up, ESD paths, and test access. Correlation retains netlists, models, scripts, tool versions, raw results, lab conditions, calibration status, and explanations for outliers. This evidence turns a nominal design into a reproducible component that can be signed off across device, circuit, package, firmware, and system teams. Corner selection should follow sensitivity rather than blindly combining labels. Deterministic sweeps expose monotonic trends, targeted Monte Carlo analysis estimates distribution tails, and importance sampling can explore rare failures. Reviewers should distinguish model uncertainty from manufacturing variation and avoid claiming yield from too few samples. The interface contract must state what happens outside normal operation. Open and short terminals, reverse polarity, hot plug, disabled bias, floating control pins, clock loss, thermal shutdown, current limiting, and repeated fault cycling often determine field reliability even though they are absent from the nominal transfer function. Dynamic behavior deserves the same attention as steady state. Settling, overshoot, ringing, slew, recovery from saturation, mode transitions, and interaction with external poles can violate a system limit long before a DC endpoint does. Time-domain tests should include realistic edge rates and source impedance. Noise should be referred to the signal or supply point that matters to the application and integrated only over a stated bandwidth. Thermal, flicker, quantization, switching, reference, substrate, and electromagnetic contributions may combine differently across modes, so a single spot-noise number rarely completes the specification. Power and thermal claims should include quiescent, active, transient, and fault states. Average efficiency can hide localized current density or hot spots; electrothermal simulation and temperature-aware device models connect electrical stress to lifetime, drift, and protection thresholds. Physical design must preserve the assumptions behind the schematic. Symmetry, common-centroid placement, dummies, shielding, guard rings, Kelvin sensing, wide current paths, via arrays, controlled coupling, and quiet reference routing are selected according to the dominant error rather than applied as decoration. Production test strategy is part of design. Trim range, observability, loopback modes, built-in self-test, boundary conditions, test time, and instrument uncertainty determine which specifications can be guaranteed economically. Characterization across wafers and lots should feed model and guard-band updates. System telemetry can extend laboratory correlation into deployed products. Error counters, calibration codes, temperatures, supply monitors, fault flags, margin measurements, and performance events help distinguish random failures from systematic drift without exposing sensitive implementation details. A useful comparison normalizes alternatives at equal output requirement and environment. Peak headline values can be misleading when bandwidth, drive, voltage, area, cooling, external components, calibration, or reliability differs; the decision record should name the workload and weighting used. Cross-functional review should trace each requirement from physical mechanism through circuit behavior to application impact. That trace prevents duplicated margin, exposes assumptions that span ownership boundaries, and makes later process or package substitutions safer. Corner selection should follow sensitivity rather than blindly combining labels. Deterministic sweeps expose monotonic trends, targeted Monte Carlo analysis estimates distribution tails, and importance sampling can explore rare failures. Reviewers should distinguish model uncertainty from manufacturing variation and avoid claiming yield from too few samples. The interface contract must state what happens outside normal operation. Open and short terminals, reverse polarity, hot plug, disabled bias, floating control pins, clock loss, thermal shutdown, current limiting, and repeated fault cycling often determine field reliability even though they are absent from the nominal transfer function.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="-apple-system,Segoe UI,Roboto,sans-serif">
  <rect x="0" y="0" width="760" height="470" fill="#0d1117"/>
  <text x="380" y="28" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">Graph Neural Networks — Learning on Structured Data</text>
  <text x="380" y="48" fill="#8b98a5" font-size="12" text-anchor="middle">aggregate neighbor features, update node embeddings — learn representations that respect graph topology</text>

  <!-- === TOP: Message passing mechanism === -->
  <rect x="25" y="62" width="710" height="145" rx="6" fill="#080d14" stroke="#233043" stroke-width="1.2"/>
  <text x="380" y="82" fill="#e6edf3" font-size="11" text-anchor="middle" font-weight="600">Message Passing (one GNN layer)</text>

  <!-- Graph visualization -->
  <circle cx="100" cy="135" r="16" fill="#0f1a2a" stroke="#60a5fa" stroke-width="1.2"/>
  <text x="100" y="139" fill="#93c5fd" font-size="8" text-anchor="middle">v</text>

  <circle cx="60" cy="100" r="12" fill="#14261f" stroke="#34d399" stroke-width="0.8"/>
  <text x="60" y="104" fill="#6ee7b7" font-size="7" text-anchor="middle">u₁</text>
  <circle cx="140" cy="100" r="12" fill="#14261f" stroke="#34d399" stroke-width="0.8"/>
  <text x="140" y="104" fill="#6ee7b7" font-size="7" text-anchor="middle">u₂</text>
  <circle cx="60" cy="170" r="12" fill="#14261f" stroke="#34d399" stroke-width="0.8"/>
  <text x="60" y="174" fill="#6ee7b7" font-size="7" text-anchor="middle">u₃</text>
  <circle cx="140" cy="170" r="12" fill="#14261f" stroke="#34d399" stroke-width="0.8"/>
  <text x="140" y="174" fill="#6ee7b7" font-size="7" text-anchor="middle">u₄</text>

  <!-- Edges -->
  <line x1="72" y1="107" x2="88" y2="124" stroke="#34d399" stroke-width="0.8"/>
  <line x1="132" y1="107" x2="112" y2="124" stroke="#34d399" stroke-width="0.8"/>
  <line x1="72" y1="163" x2="88" y2="146" stroke="#34d399" stroke-width="0.8"/>
  <line x1="132" y1="163" x2="112" y2="146" stroke="#34d399" stroke-width="0.8"/>

  <!-- Message arrows -->
  <path d="M72,107 L88,124" fill="none" stroke="#fbbf24" stroke-width="0.6" stroke-dasharray="2,1"/>
  <path d="M132,107 L112,124" fill="none" stroke="#fbbf24" stroke-width="0.6" stroke-dasharray="2,1"/>
  <path d="M72,163 L88,146" fill="none" stroke="#fbbf24" stroke-width="0.6" stroke-dasharray="2,1"/>
  <path d="M132,163 L112,146" fill="none" stroke="#fbbf24" stroke-width="0.6" stroke-dasharray="2,1"/>

  <!-- Steps on right -->
  <text x="200" y="105" fill="#fbbf24" font-size="9" font-weight="600">1. Message</text>
  <text x="280" y="105" fill="#8b98a5" font-size="8.5">m_uv = MLP(h_u, h_v, e_uv)</text>

  <text x="200" y="130" fill="#34d399" font-size="9" font-weight="600">2. Aggregate</text>
  <text x="290" y="130" fill="#8b98a5" font-size="8.5">M_v = Σ or max or mean over N(v)</text>

  <text x="200" y="155" fill="#60a5fa" font-size="9" font-weight="600">3. Update</text>
  <text x="270" y="155" fill="#8b98a5" font-size="8.5">h_v' = GRU(h_v, M_v) or MLP(h_v ‖ M_v)</text>

  <text x="200" y="180" fill="#c4b5fd" font-size="9" font-weight="600">4. Readout (graph-level)</text>
  <text x="355" y="180" fill="#8b98a5" font-size="8.5">h_G = pool({h_v | v ∈ V})</text>

  <!-- k-hop note -->
  <text x="560" y="120" fill="#6b7684" font-size="8.5">After k layers, each node has</text>
  <text x="560" y="135" fill="#6b7684" font-size="8.5">information from its k-hop</text>
  <text x="560" y="150" fill="#6b7684" font-size="8.5">neighborhood (receptive field)</text>
  <text x="560" y="170" fill="#fbbf24" font-size="8.5">Typical: 2-5 layers (over-smoothing)</text>

  <!-- === MIDDLE LEFT: GNN variants === -->
  <rect x="25" y="215" width="350" height="112" rx="6" fill="#0b1220" stroke="#233043" stroke-width="1"/>
  <text x="200" y="233" fill="#e6edf3" font-size="10" text-anchor="middle" font-weight="600">GNN Architectures</text>

  <text x="45" y="253" fill="#60a5fa" font-size="8.5" font-weight="600">GCN (Kipf 2017):</text>
  <text x="160" y="253" fill="#8b98a5" font-size="8.5">mean aggregate, symmetric norm</text>
  <text x="45" y="269" fill="#34d399" font-size="8.5" font-weight="600">GAT:</text>
  <text x="80" y="269" fill="#8b98a5" font-size="8.5">attention-weighted aggregation</text>
  <text x="45" y="285" fill="#fbbf24" font-size="8.5" font-weight="600">GraphSAGE:</text>
  <text x="125" y="285" fill="#8b98a5" font-size="8.5">sample neighbors, inductive (new nodes)</text>
  <text x="45" y="301" fill="#c4b5fd" font-size="8.5" font-weight="600">GIN:</text>
  <text x="75" y="301" fill="#8b98a5" font-size="8.5">sum aggregate (maximally expressive)</text>
  <text x="45" y="317" fill="#f87171" font-size="8.5" font-weight="600">MPNN:</text>
  <text x="85" y="317" fill="#8b98a5" font-size="8.5">general message passing framework</text>

  <!-- === MIDDLE RIGHT: Applications === -->
  <rect x="390" y="215" width="345" height="112" rx="6" fill="#0b1220" stroke="#233043" stroke-width="1"/>
  <text x="562" y="233" fill="#e6edf3" font-size="10" text-anchor="middle" font-weight="600">Applications</text>

  <text x="410" y="253" fill="#60a5fa" font-size="8.5" font-weight="600">Drug discovery:</text>
  <text x="510" y="253" fill="#8b98a5" font-size="8.5">molecule graphs → property prediction</text>
  <text x="410" y="269" fill="#34d399" font-size="8.5" font-weight="600">Social networks:</text>
  <text x="510" y="269" fill="#8b98a5" font-size="8.5">user graphs → recommendation, fraud</text>
  <text x="410" y="285" fill="#fbbf24" font-size="8.5" font-weight="600">EDA / chip design:</text>
  <text x="530" y="285" fill="#8b98a5" font-size="8.5">netlist graphs → timing, congestion</text>
  <text x="410" y="301" fill="#c4b5fd" font-size="8.5" font-weight="600">Physics simulation:</text>
  <text x="530" y="301" fill="#8b98a5" font-size="8.5">particle graphs → dynamics (DeepMind)</text>
  <text x="410" y="317" fill="#f87171" font-size="8.5" font-weight="600">Knowledge graphs:</text>
  <text x="525" y="317" fill="#8b98a5" font-size="8.5">entity/relation → link prediction</text>

  <!-- === BOTTOM: Libraries === -->
  <rect x="25" y="337" width="710" height="65" rx="5" fill="#0b1220" stroke="#233043" stroke-width="1"/>
  <text x="380" y="355" fill="#e6edf3" font-size="10" text-anchor="middle" font-weight="600">Frameworks</text>

  <text x="120" y="377" fill="#c4b5fd" font-size="9.5" text-anchor="middle" font-weight="600">PyG (PyTorch Geo)</text>
  <text x="120" y="391" fill="#8b98a5" font-size="8" text-anchor="middle">most popular, modular</text>

  <text x="300" y="377" fill="#fbbf24" font-size="9.5" text-anchor="middle" font-weight="600">DGL</text>
  <text x="300" y="391" fill="#8b98a5" font-size="8" text-anchor="middle">AWS, scalable, multi-backend</text>

  <text x="470" y="377" fill="#34d399" font-size="9.5" text-anchor="middle" font-weight="600">JAX-MD / e3nn</text>
  <text x="470" y="391" fill="#8b98a5" font-size="8" text-anchor="middle">equivariant, physics</text>

  <text x="630" y="377" fill="#60a5fa" font-size="9.5" text-anchor="middle" font-weight="600">NetworkX</text>
  <text x="630" y="391" fill="#8b98a5" font-size="8" text-anchor="middle">classical graph algos</text>

  <!-- Key insight -->
  <rect x="25" y="410" width="710" height="22" rx="3" fill="#0b1220" stroke="#233043" stroke-width="0.8"/>
  <text x="380" y="425" fill="#fbbf24" font-size="9" text-anchor="middle">GNNs generalize CNNs to irregular structure: a CNN is just a GNN on a regular pixel grid with fixed neighborhood.</text>

  <text x="380" y="460" fill="#6b7684" font-size="11" text-anchor="middle">Whenever your data has explicit relationships (edges), a GNN can exploit that structure — molecules, chips, networks, physics.</text>
</svg>

Connection to CFS platform. Use CFS AI, accelerator, memory, networking, serving, sensor, robotics, and system simulators with linked glossary topics to connect application behavior to measurable hardware and deployment trade-offs.