Banded-Router Architecture Overview

• Banded-Router Architecture is a design strategy that partitions routing into bounded subdomains such as layers, banks, or tiers to simplify control and enhance performance.
• It decomposes traffic spatially, temporally, or semantically across various domains—from packet switches and photonic fabrics to FPGA, quantum, and LLM systems.
• Key implementations demonstrate tradeoffs in energy, delay, and throughput, using methods like load balancing, passive splitting, and contextual bandit models for routing decisions.

Banded-router architecture, used here as an Editor’s term for a heterogeneous family of routing designs, denotes architectures that partition routing over bounded sets, layers, banks, tiers, or independently controlled subfabrics rather than over a single monolithic decision surface. Across packet switches, photonic fabrics, FPGA NoCs, quantum repeaters, and LLM routing systems, the recurring architectural move is structural decomposition: traffic is split spatially, temporally, semantically, or by candidate set, and each partition is handled by a restricted routing policy or switching resource. The literature does not employ the term uniformly. Several papers explicitly treat “banded” only as a conceptual analogy, while others instead formalize related notions such as load-balanced middle stages, split-parallel switches, hybrid two-layer routers, left/right memory banks, or tiered semantic and safety routing (Andrews et al., 2013, Keslassy et al., 11 Feb 2026, Ezhumalai et al., 2010, Chen et al., 22 Mar 2026).

1. Terminological scope and conceptual boundaries

The term is not a standardized router taxonomy across the cited literature. In internet switching and optics, the closest formal notions are structured middle stages, permutation-matrix fabrics, and passive spatial division. In FPGA and quantum systems, the comparable constructs are two-layer packet/circuit separation and banked local connectivity. In LLM systems, the analogous idea is a bounded or tiered dispatch policy over models, agents, pools, or safety classes rather than a single flat classifier (Keslassy et al., 11 Feb 2026, Ezhumalai et al., 2010, Lee et al., 2020, Chen et al., 22 Mar 2026).

A common misconception is to conflate banded routing with bandit-based routing. OrcaRouter is explicitly a production-oriented LLM router built around a multi-arm contextual bandit using LinUCB, not a “banded” fabric in the switching-theoretic sense. It formulates routing over a model pool $\mathcal{A}=\{1,\dots,K\}$, observes a request context $x_t\in\mathbb{R}^d$, selects an arm $a_t$, and maximizes cumulative reward $\max \sum_{t=1}^{T} r_t$, where the reward combines quality, normalized cost, normalized latency, and operational penalties (Bao et al., 29 May 2026).

This terminological ambiguity is explicit in the multi-agent literature. The TCAndon-Router paper states that, if by “Banded-Router Architecture” one means a router that selects from a bounded set of candidates rather than a single expert, then its reasoning-first subset router fits that idea; however, the paper’s own formalism is that of set-valued agent routing with collaborative execution and refinement rather than a canonical banded-router taxonomy (Zhao et al., 8 Jan 2026).

2. Core structural patterns

Across the papers, the unifying design principle is constrained decomposition. Instead of fully dynamic all-to-all routing with one universal scheduler or one universal predictor, the system is divided into intermediate stages, banks, pools, or bands with restricted interaction. This suggests a cross-domain pattern: routing complexity is reduced by replacing unrestricted choice with structured locality, staged control, or bounded candidate sets.

Domain	Partitioning primitive	Reported purpose
Load-balanced packet router	Active middle-stage nodes in an $n\times m\times n$ fabric	Avoid centralized scheduling; trade energy against queue and delay
Split-parallel internet router	$H$ independent smaller switches via passive fiber split	Petabit/sec scaling within a single integrated package
FPGA NoC router	Packet-switched P-layer plus circuit-switched C-layer	Reduce local control and serialization overhead
Quantum repeater router	Left and right memory banks plus local switchboard	Improve fidelity as multiplexing depth increases
LLM and agent routers	Candidate models, agents, pools, and safety tiers	Balance quality, cost, latency, safety, and adaptation

In the load-balanced router, each packet is sent from input $i$ to a randomly chosen middle-stage node and then forwarded to output $k$, using two input-buffered meshes and FIFO service. In the split-parallel internet router, incoming fibers are spatially divided across $H$ independent switches without intermediate OEO conversions for balancing. In the FPGA design, nearby local traffic is separated from general NoC traffic by a distinct circuit-switched layer. In the quantum router, left-bank and right-bank memories are paired through a local photonic switchboard. In LLM systems, the same architectural role is played by semantic bands, safety tiers, token-budget pools, or bounded candidate subsets (Andrews et al., 2013, Keslassy et al., 11 Feb 2026, Ezhumalai et al., 2010, Lee et al., 2020, Chen et al., 22 Mar 2026).

3. Classical packet, optical, and in-package realizations

The most explicit switching-theoretic instantiation is the Load-Balanced Router. It is a three-stage architecture with $n$ input ports, $x_t\in\mathbb{R}^d$0 intermediate nodes, and $x_t\in\mathbb{R}^d$1 output ports. Traffic from input $x_t\in\mathbb{R}^d$2 to output $x_t\in\mathbb{R}^d$3 is assigned to a middle node $x_t\in\mathbb{R}^d$4 chosen uniformly at random among the active middle-stage nodes; the internal link rates are $x_t\in\mathbb{R}^d$5 from input to middle and $x_t\in\mathbb{R}^d$6 from middle to output. Its energy-proportionality lever is that only $x_t\in\mathbb{R}^d$7 middle-stage nodes need be active, with energy consumption $x_t\in\mathbb{R}^d$8. The paper states that the router is not overloaded if $x_t\in\mathbb{R}^d$9, and derives exponential tail bounds for queue size and delay using Chernoff bounds, thereby formalizing an energy–delay tradeoff (Andrews et al., 2013).

A more recent large-scale variant is the Split-Parallel Switch (SPS) architecture for internet routers. SPS decomposes a router into $a_t$0 parallel smaller switches, each handling about $a_t$1 of total traffic, and performs a passive spatial split of fibers rather than per-packet electronic load balancing. The reference design uses $a_t$2 fiber-ribbon arrays, each with $a_t$3 fibers, $a_t$4 WDM channels per fiber, and rate $a_t$5 Gb/s per wavelength, yielding $a_t$6 Tb/s per direction and $a_t$7 Pb/s total I/O. With $a_t$8, each smaller switch handles $a_t$9 Tb/s of combined input and output memory I/O. Each smaller switch is an HBM-based shared-memory switch using $\max \sum_{t=1}^{T} r_t$0 HBM4 stacks and $\max \sum_{t=1}^{T} r_t$1 parallel memory channels, and its Parallel Frame Interleaving (PFI) algorithm aggregates packets into $\max \sum_{t=1}^{T} r_t$2 KB batches and $\max \sum_{t=1}^{T} r_t$3 KB frames to sustain peak HBM bandwidth (Keslassy et al., 11 Feb 2026).

The photonic-plasmonic literature provides a different physical realization of structured routing. The hybrid photonic-plasmonic non-blocking broadband 5×5 router is built as a 5-port spatial switching fabric using eight 2×2 hybrid photonic-plasmonic switches. The architecture is permutation-matrix-style and non-blocking: each 2×2 switch can be set to BAR or CROSS, and the eight-switch arrangement implements all non-self 5×5 routes while avoiding waveguide crossings. The router is explicitly non-resonant, avoiding wavelength-selective resonance and thermal tuning, and the paper reports an average 3 dB spectral bandwidth of 106 nm, approximately spanning 1.49 µm to 1.62 µm, with support for CWDM and DWDM operation. The important nuance is that the device carries multiple pre-multiplexed wavelengths together; it does not route individual wavelengths separately as a wavelength-selective resonator router would (Sun et al., 2017).

4. Locality-aware hybridization in FPGA and quantum routers

The high-performance hybrid two-layer router in the MoCReS framework makes the locality principle explicit. Its central claim is that strictly packet-switched routers on FPGAs incur unnecessary control and serialization overhead for IP cores attached to the same router. Baseline MoCReS uses a req/grant protocol with round-robin arbitration and requires at least 6 cycles for input data to appear at the downstream router or local IP. The proposed router therefore combines a packet-switched layer (P-layer) for communication through the mesh with a time-multiplexed circuit-switched layer (C-layer) for directly connected local IPs. The P-layer retains multi-clock virtual cut-through and deadlock-free XY routing; the C-layer provides direct scheduled communication, uses a 32-bit bus width rather than the P-layer’s 8-bit channels, and removes local-to-local traffic from the packet-switched fabric for selected IPs. The reported result is an average improvement of 20.4 percent in NoC bandwidth, with a maximum of 24 percent compared to a traditional NoC (Ezhumalai et al., 2010).

The paper is explicit that this is not a classical banded router in a standard crossbar taxonomy. Its affinity to a banded architecture is conceptual: traffic is partitioned into a high-throughput direct/local region and a network-routed region, and the router becomes a hybrid multi-structure fabric whose arbitration logic and cross-point matrix are simplified by that separation. The price is architectural constraint: more C-layer ports increase schedule memory, can reduce clock rate, and require synchronous IPs because the C-layer is unbuffered (Ezhumalai et al., 2010).

The quantum networking literature arrives at a comparable structure for different physical reasons. The quantum router architecture for high-fidelity entanglement flows divides the qubit registers at a repeater node into left-facing and right-facing banks, each connected optically only to its corresponding network side, while a local photonic switchboard provides all-to-all connectivity across the bank boundary. In the concrete implementation, each register is an NV center with an electron spin as entanglement broker and a nuclear spin as storage. Successful left-side and right-side distant entanglements are stored, then greedily paired in first-in-first-out order and connected through local entanglement generation and Bell-state measurements. The main reported consequence is channel-loss-invariant entanglement fidelity: as link loss increases, fidelity no longer degrades because the dominant waiting process shifts from slow distant attempts to fast local pairing (Lee et al., 2020).

This design also exhibits a characteristic banded tradeoff. The router rate is somewhat lower than the routerless multiplexed repeater for small multiplexing depth $\max \sum_{t=1}^{T} r_t$4, because left/right success mismatches reduce immediate pairing efficiency; however, the gap closes as $\max \sum_{t=1}^{T} r_t$5 increases, while fidelity improves substantially. For short links, the router’s extra local operations can dominate and reduce fidelity relative to the standard repeater; for lossier channels, the banked switchboard architecture becomes advantageous because it shortens storage time (Lee et al., 2020).

5. Semantic and expert-selection analogues in LLM systems

In LLM routing, the “banded” analogy shifts from physical switch fabrics to policy layers. OrcaRouter formalizes request dispatch as a contextual bandit over a fixed model pool. Each request is encoded using a hybrid feature representation: a 40-dimensional handcrafted lexical vector together with a frozen all-MiniLM-L6-v2 384-dimensional sentence embedding, producing the augmented context vector

$\max \sum_{t=1}^{T} r_t$6

Offline, the system evaluates every candidate model on a curated training set and fits one ridge regressor per arm,

$\max \sum_{t=1}^{T} r_t$7

Online, it may remain frozen with $\max \sum_{t=1}^{T} r_t$8 or continue updating via LinUCB with score

$\max \sum_{t=1}^{T} r_t$9

The warmup set in the experiments contains 5,000 filtered RouterBench prompts evaluated across 10 candidate models, after removing 1,135 near-duplicate prompts whose MiniLM cosine similarity to any RouterArena prompt exceeds 0.85. At the time of submission, OrcaRouter-Adaptive ranked second on the public RouterArena leaderboard with arena score 72.08, accuracy 75.54%, and cost USD 1.00 per 1,000 queries (Bao et al., 29 May 2026).

The TCAndon-Router (TCAR) is structurally closer to a bounded-candidate or subset router than to a flat classifier. It defines routing as

$n\times m\times n$0

where $n\times m\times n$1 is a natural-language reasoning chain and $n\times m\times n$2 is a selected set of agents rather than a single label. The prompt incorporates the instruction, the query, and textual capability descriptions of all agents. After routing, each selected agent independently answers the query, and a dedicated Refining Agent aggregates and rewrites those outputs into a single final response. This design is intended for enterprise settings with overlapping agent responsibilities and dynamic agent onboarding, where static single-label routing is brittle. The paper reports that reasoning improves performance across all evaluated datasets, and that on QCloud the TCAR(4B) score is 93.98. It also notes that average reasoning length is under 100 tokens, average selected agents is 1.37, and latency remains under 1 second on GPUs for the 4B model (Zhao et al., 8 Jan 2026).

The Workload–Router–Pool (WRP) architecture generalizes the LLM-router view even further. It defines a deployment as a triple $n\times m\times n$3, where workload characterizes request distributions, router maps requests to a $n\times m\times n$4 tuple, and pool characterizes the inference substrate. Within this framework, routing already appears in explicitly tiered forms: SafetyL1 as a binary safe/unsafe gate, SafetyL2 as a nine-class hazard classifier, token-budget pool routing for short-context versus long-context requests, semantic cache routing, online user-feedback-driven routing adaptation, and RL-style routing for agentic workloads. The paper’s own conclusion is that the strongest open questions lie at the Router × Pool and Pool × Router intersections, which is precisely where a multi-band, hierarchical, staged router interpretation becomes most natural (Chen et al., 22 Mar 2026).

6. Tradeoffs, stability, and unresolved questions

A banded-router architecture typically exchanges unrestricted optimization for deployable structure. In the load-balanced router, fewer active middle-stage nodes save energy but worsen queueing and delay; the benefit is energy proportionality, while the cost is a measurable degradation in tail behavior. In SPS, passive spatial division avoids OEO-heavy fine-grained balancing, but a coarse split can overload one smaller switch under skewed traffic even when the global traffic matrix is admissible. The paper nevertheless reports very small loss rates in evaluation, including under 0.013% on resized Abilene and GEANT workloads, below 0.008% on a CAIDA trace, below 0.05% on synthetic WAN traffic, and under 0.35% on cross-datacenter AI workloads. Its scaling bottleneck is identified as power: the estimate is about 794 W per HBM switch and approximately 12.7 kW for 16 switches (Andrews et al., 2013, Keslassy et al., 11 Feb 2026).

Stability and sensitivity recur in semantic routers as well. OrcaRouter notes that paraphrase robustness is a major issue because small embedding shifts can flip the top-ranked arm. To reduce pick-flip rate, it proposes a tie-breaking rule: if the top two arms are within $n\times m\times n$5, choose the arm with the higher mean training reward. The paper reports that this reduces pick-flip rate from 16.7% to 2.4%, at the cost of about 0.86 Arena points (Bao et al., 29 May 2026).

In TCAR, the principal limitations are dependence on agent description quality, long-tail and rare-domain issues, reasoning–prediction mismatch, and instruction-format sensitivity. In the FPGA router, increasing the number of C-layer ports enlarges schedule memory and harms frequency and signal integrity. In the quantum router, short links can negate the fidelity advantage because the architecture requires extra local operations per delivered EPR pair. In WRP, the unresolved questions are not primarily about model choice in isolation, but about the coupling between routing policy and pool state, including pool-aware cascading, KV-cache retention directives, and closed-loop self-adaptation (Zhao et al., 8 Jan 2026, Ezhumalai et al., 2010, Lee et al., 2020, Chen et al., 22 Mar 2026).

Taken together, these works suggest that banded-router architecture is best understood not as a single canonical fabric, but as a recurrent systems strategy: divide the routing problem into constrained subdomains, accept bounded interaction between them, and use that structure to make scale, energy, latency, reliability, or interpretability tractable. The specific band—middle-stage node, fiber subset, local circuit layer, memory bank, model pool, safety tier, or agent subset—varies by domain, but the architectural logic is consistent.