Edge Efficiency Index (EEI): Metrics & Methods

• Edge Efficiency Index (EEI) is a multidimensional metric that evaluates edge performance by normalizing useful work against resource constraints such as latency, energy, and cost.
• Different formulations of EEI apply to embedded vision, edge–cloud energy benchmarking, 5G cost-weighted throughput, and topological efficiency in network graphs.
• Practical applications include real-time embedded computing, autonomous logistics, and network optimization, guiding efficient use of resources in constrained environments.

Edge Efficiency Index (EEI) denotes a class of edge-oriented efficiency measures rather than a single canonical scalar. Across recent work, the term is either explicit or implicit in four distinct but related senses: resource-normalized efficiency for embedded visual feature extraction, confidence-aware energy efficiency over the edge–cloud continuum, cost-weighted throughput efficiency for 5G systems with X-Haul and edge caching, and topological edge criticality measured through edge proximity (Yao et al., 24 Apr 2025, Jahnke et al., 10 Mar 2025, Yan et al., 2016, Banerjee et al., 2014). In all four settings, the common objective is to quantify useful edge-side capability relative to a constrained denominator such as latency, FLOPs, parameter count, energy, deployment cost, communication payload, or structural path efficiency.

1. Conceptual scope and formal families

The literature does not supply a universally standardized EEI. In "EdgePoint2" (Yao et al., 24 Apr 2025) and "GMB-ECC" (Jahnke et al., 10 Mar 2025), the authors do not define a literal metric called “Edge Efficiency Index (EEI)” explicitly. Instead, both works provide nearly complete ingredients for such an index. By contrast, "Economical Energy Efficiency $E^3$" (Yan et al., 2016) defines a cost-aware throughput-over-energy ratio that directly functions as a template for an edge-specific EEI, while "edge proximity" $\mathcal{P}_e$ (Banerjee et al., 2014) provides an edge-centric topological measure that can serve directly as an edge efficiency indicator.

Research context	Representative quantity	Efficiency interpretation
Embedded local features	$\text{EEI}_1 = \frac{A}{T \cdot S}$, $\text{EEI}_2 = \frac{A}{F \cdot S}$, $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$	Accuracy per latency, model size, compute, or communication payload
Edge–cloud energy benchmarking	$$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$	Proximity to optimal normalized energy efficiency, with uncertainty
5G edge systems	$$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$	Throughput over cost-weighted energy
Network topology	$\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$	Edge centrality in maintaining short edge–edge paths

This suggests that EEI is best understood as a domain-specific normalized efficiency functional. The numerator measures useful work or utility at the edge; the denominator or complement term captures the dominant constraint of the application domain.

2. Resource-normalized EEI in embedded visual computing

In "EdgePoint2: Compact Descriptors for Superior Efficiency and Accuracy" (Yao et al., 24 Apr 2025), efficiency on edge and embedded devices is defined through a coupled evaluation of compute footprint, runtime on actual edge hardware, descriptor dimensionality, and downstream task accuracy. The paper explicitly couples efficiency with number of parameters, GFLOPs per image, FPS on NVIDIA Jetson Orin‑NX GPU using TensorRT, FPS on the Orin‑NX ARM CPU using OpenVINO, descriptor dimensionality $32/48/64$, and accuracy on HPatches, MegaDepth‑1500, ScanNet‑1500, IMC2022, FM‑Bench, Aachen Day‑Night v1.1, and InLoc. The operative notion is “accuracy at a given resource usage,” or “resource usage for a given accuracy.”

The architectural strategy is deliberately minimalist. EdgePoint2 uses a very small backbone with simple convolutions only: two standard convolutions, two ResNet blocks, pooling, and ReLU activations, with no deformable convolutions, attention, or other heavy operations. It also applies aggressive spatial downsampling, then separates detection and description heads by scale: the detection head uses a $\frac{H}{2} \times \frac{W}{2}$ feature map, while the description head uses a $\mathcal{P}_e$0 feature map with pyramid aggregation, a $\mathcal{P}_e$1 convolution, a group convolution, and a final convolution that produces descriptor dimension $\mathcal{P}_e$2. This implements an explicit accuracy–efficiency trade-off.

Descriptor compactness is achieved through low-rank distillation. All descriptors are L2-normalized, and the training objective preserves the teacher’s cosine similarity matrix while compressing descriptor dimensionality. The paper introduces Orthogonal Procrustes loss and similarity loss as a general approach for hypersphere embedding distillation tasks. The total loss is

$\mathcal{P}_e$3

with $\mathcal{P}_e$4, $\mathcal{P}_e$5, and $\mathcal{P}_e$6. The practical consequence is that 32-, 48-, and 64-dimensional descriptors can match or beat many 64/128-dimensional baselines, directly reducing descriptor storage per keypoint, matching cost, and communication cost in multi-robot or bandwidth-limited settings.

The model family contains 14 sub-models: T32, T48, S32, S48, S64, M32, M48, M64, L32, L48, L64, E32, E48, and E64. Runtime measurements on Jetson Orin‑NX at input $\mathcal{P}_e$7 and 1024 keypoints illustrate the resulting operating points.

Model	Resources	FPS
EdgePoint2‑T32	Dim $\mathcal{P}_e$8, MP $\mathcal{P}_e$9 M, GFLOPs $\text{EEI}_1 = \frac{A}{T \cdot S}$0	508.7 GPU, 40.84 CPU
EdgePoint2‑E64	Dim $\text{EEI}_1 = \frac{A}{T \cdot S}$1, MP $\text{EEI}_1 = \frac{A}{T \cdot S}$2 M, GFLOPs $\text{EEI}_1 = \frac{A}{T \cdot S}$3	375.7 GPU, 16.92 CPU
XFeat	$\text{EEI}_1 = \frac{A}{T \cdot S}$4D, $\text{EEI}_1 = \frac{A}{T \cdot S}$5 M params, $\text{EEI}_1 = \frac{A}{T \cdot S}$6 GFLOPs	277.4 GPU, 35.33 CPU
ALIKED‑N32	$\text{EEI}_1 = \frac{A}{T \cdot S}$7D, $\text{EEI}_1 = \frac{A}{T \cdot S}$8 M params, $\text{EEI}_1 = \frac{A}{T \cdot S}$9 GFLOPs	29.7 GPU, 1.12 CPU

The paper does not define EEI explicitly, but it proposes several formulations consistent with its methodology: $\text{EEI}_2 = \frac{A}{F \cdot S}$0 It also proposes a general form,

$\text{EEI}_2 = \frac{A}{F \cdot S}$1

and a model-level instantiation,

$\text{EEI}_2 = \frac{A}{F \cdot S}$2

Here $\text{EEI}_2 = \frac{A}{F \cdot S}$3 is a task accuracy metric, $\text{EEI}_2 = \frac{A}{F \cdot S}$4 is latency, $\text{EEI}_2 = \frac{A}{F \cdot S}$5 is model size, $\text{EEI}_2 = \frac{A}{F \cdot S}$6 is compute cost, $\text{EEI}_2 = \frac{A}{F \cdot S}$7 is descriptor dimension, and $\text{EEI}_2 = \frac{A}{F \cdot S}$8 is a normalized communication cost term. In this formulation, EEI rewards more task performance per byte, per FLOP, and per unit latency.

3. Confidence-aware EEI across the edge–cloud continuum

"GMB-ECC: Guided Measuring and Benchmarking of the Edge Cloud Continuum" (Jahnke et al., 10 Mar 2025) frames EEI as a normalized, uncertainty-aware energy-efficiency measure over heterogeneous systems. The framework represents the edge–cloud continuum as a weighted directed acyclic graph for each state $\text{EEI}_2 = \frac{A}{F \cdot S}$9,

$\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$0

where $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$1 are components, $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$2 are dependency edges, and $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$3 are utilization weights. The framework distinguishes measurable components $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$4, which are leaf nodes with physically measurable units such as CPUs, GPUs, NICs, sensors, and storage, from composite components $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$5, which aggregate subsystems such as an autonomous vehicle, an edge server, or an entire pipeline.

Each node $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$6 in state $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$7 is assigned a normalized efficiency function

$\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$8

mapping utilization to normalized efficiency. For composite components $\text{EEI}_{\text{comm}} = \frac{A}{b D K \cdot T}$9, efficiency is aggregated as

$$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$0

and variance is propagated as

$$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$1

This yields both normalized efficiency and a precision parameter for any composite edge or edge–cloud entity.

Time aggregation over a state set $$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$2 produces

$$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$3

For each component, the optimal utilization is

$$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$4

and the central KPI is the efficiency gap

$$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$5

GMB-ECC then classifies components as “well‑tuned,” “partially optimized,” or “misconfigured” using thresholds $$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$6 and $$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$7.

A natural EEI follows directly: $$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$8 This yields a dimensionless efficiency index in $$\text{EEI}_{k,S^\ast} = 1 - \Delta \eta_{max,k,S^\ast}$$9 that is simultaneously efficiency-aware and confidence-aware. The same work also distinguishes

$$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$0

thereby separating pure normalized efficiency from distance-to-optimum with uncertainty.

The autonomous intra-logistics use case shows how such an EEI operates in practice. The scenario includes autonomous vehicles in warehouses, each with LiDAR, cameras, GPS, an NVIDIA Jetson AGX Xavier, and a 5G modem (Quectel RM500Q‑GL), plus edge servers and cloud analytics on Dell PowerEdge R740. Over 7 days, the framework identifies processor and communication modules as optimization targets. Two optimizations are reported: Dynamic Voltage and Frequency Scaling reduces processor energy from 25 W to 20 W, and adaptive transmission intervals plus LZ4 compression reduce modem energy from 4 W to 3 W. Overall vehicle consumption drops from 50 W to 44 W, decision latency remains unchanged, and measurement precision is tuned so error margin is less than 5%. In EEI terms, the composite efficiency increases and the efficiency gap decreases.

4. Cost-weighted EEI for edge radio, X-Haul, and caching

"Economical Energy Efficiency $$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$1: An Advanced Performance Metric for 5G Systems" (Yan et al., 2016) defines a cost-aware metric for systems in which X-Haul and edge caching are first-class components. Its purpose is to evaluate jointly the useful throughput delivered by a 5G RAN, the energy consumption of its radio and edge infrastructure, and the economic cost of that infrastructure and its operation. The formal definition is

$$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$2

where $$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$3 is effective throughput of UE $$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$4, $$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$5 is a priority weight, $$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$6 is dynamic power of BS $$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$7, $$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$8 is static power, and $$E^3 = \frac{\sum_{k \in \mathcal{K}} \alpha_k R_k}{\sum_{n \in \mathcal{N}} (P_{Tn} + P_{0n} C_n)}$$9 is a dimensionless cost coefficient. The coefficient is defined by

$\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$0

with $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$1 the average cost per unit area for BSs of the same type as $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$2 and $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$3 a benchmark cost per unit area.

The significance of $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$4 is that SE ignores energy and cost, EE ignores cost, and cost-only metrics ignore energy. This is especially consequential in architectures such as C‑RAN, H‑CRAN, and F‑RAN, where expensive optical fiber X-Haul, cheaper wireless X-Haul, and different amounts of edge cache can yield similar SE or EE but radically different CAPEX/OPEX.

The paper’s qualitative findings are directly relevant to EEI design. Increasing X-Haul capacity beyond what traffic actually needs can cause $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$5 to decrease because cost rises while throughput and energy do not improve proportionally. With fixed X-Haul capacity, increasing cache size improves SE, EE, and $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$6 initially, but beyond some cache size $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$7 starts to decrease once extra cache contributes little extra hit ratio but significantly increases cost coefficient $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$8. An advanced caching strategy yields higher $\mathcal{P}_e = \frac{M-1}{\sum_{f \in E} d(e,f)}$9 than a referential caching strategy for the same cache size. Under joint X-Haul and cache variation, some combinations of moderate X-Haul and moderate cache yield higher $32/48/64$0 than large X-Haul and large cache because of cost.

A plausible implication is an edge-specific EEI with the same structure: $32/48/64$1 This adaptation restricts the numerator to traffic served through edge resources and the denominator to energy and cost attributable to edge nodes, edge caches, and associated X-Haul. The paper is explicit, however, that cost modeling remains an open problem.

5. Topological EEI and edge proximity

"Slow poisoning and destruction of networks: Edge proximity and its implications for biological and infrastructure networks" (Banerjee et al., 2014) supplies a structurally different notion of EEI: topological efficiency at the level of individual edges. The paper defines edge proximity $32/48/64$2 for an edge $32/48/64$3 of a connected graph $32/48/64$4 as

$32/48/64$5

where $32/48/64$6 and $32/48/64$7 is the shortest-path distance between edges, equivalently the shortest-path distance between the corresponding nodes in the line graph $32/48/64$8. Thus $32/48/64$9 is exactly the closeness centrality of the node corresponding to edge $\frac{H}{2} \times \frac{W}{2}$0 in $\frac{H}{2} \times \frac{W}{2}$1.

The associated global efficiency measure is

$\frac{H}{2} \times \frac{W}{2}$2

with complementary tracking of average shortest path length

$\frac{H}{2} \times \frac{W}{2}$3

and diameter

$\frac{H}{2} \times \frac{W}{2}$4

The paper repeatedly removes edges by maximum $\frac{H}{2} \times \frac{W}{2}$5, maximum edge betweenness $\frac{H}{2} \times \frac{W}{2}$6, maximum edge degree $\frac{H}{2} \times \frac{W}{2}$7, and random selection. The main finding is the “slow poisoning” effect: targeting by $\frac{H}{2} \times \frac{W}{2}$8 may keep the network connected for a long time, yet it leads to a remarkable increase in diameter and average shortest path length until the first disconnection and well beyond, and it causes notable efficiency loss in U.S. and European power grid networks.

The same metric identifies functionally important interactions across domains. In protein–protein interaction networks it highlights interactions involving essential proteins; in the directed neural network of C. elegans it pinpoints various synapses of AVEL solely associated with backward movement; in macaque brain networks it identifies connections from cortex to thalamus, frontal lobe, and temporal lobe that are starting interactions of longer information processing pathways; and in food webs it identifies root interactions that form the backbone of long food web chains. The paper also notes that edges with higher $\frac{H}{2} \times \frac{W}{2}$9 can be relevant to edge controllability under switchboard dynamics.

In this line of work, a direct EEI is simply

$\mathcal{P}_e$00

A plausible impact-based alternative is

$\mathcal{P}_e$01

with $\mathcal{P}_e$02 acting as a topological proxy for efficiency loss under removal. The paper further notes that $\mathcal{P}_e$03 becomes less informative in very small networks and has limited importance in very dense graphs.

6. Comparative interpretation, misconceptions, and limitations

A common misconception is to treat EEI as synonymous with energy efficiency. The cited work shows a broader structure. In embedded vision, EEI-like behavior combines accuracy with runtime, FLOPs, parameter count, and descriptor dimensionality (Yao et al., 24 Apr 2025). In edge–cloud benchmarking, the index is proximity to optimal normalized energy efficiency under explicit variance propagation (Jahnke et al., 10 Mar 2025). In 5G systems, the decisive addition is infrastructure cost through $\mathcal{P}_e$04 (Yan et al., 2016). In network science, the relevant quantity is not energy at all but an edge’s contribution to maintaining short, efficient paths (Banerjee et al., 2014).

A second misconception is that a single scalar can be compared across all edge domains without qualification. GMB-ECC normalizes each efficiency curve to $\mathcal{P}_e$05, so a component at $\mathcal{P}_e$06 and another component at $\mathcal{P}_e$07 are each close to their own optimum, not necessarily equal in absolute Joules per task. EdgePoint2’s proposed EEI forms are likewise task-specific, because $\mathcal{P}_e$08 may denote MHA, AUC, or localization success, and because communication-aware variants depend on descriptor dimensionality $\mathcal{P}_e$09, number of keypoints $\mathcal{P}_e$10, and bytes per dimension $\mathcal{P}_e$11. This suggests that cross-paper comparison requires attention to normalization target, operating point, and denominator semantics.

A third misconception is that accuracy or throughput alone determines edge efficiency. The 5G $\mathcal{P}_e$12 results show that overprovisioned X-Haul or excessively large caches can reduce economical efficiency even when conventional metrics saturate. EdgePoint2 shows that descriptor dimensionality is not merely a storage detail but a first-order term in communication and matching cost. GMB-ECC further shows that uncertainty matters: the term

$\mathcal{P}_e$13

inflates the efficiency gap when measurement confidence is low.

The limitations are equally domain-specific. GMB-ECC currently relies on synthetic data only, and its limitations section highlights external validity, partial observability, and dependence on manufacturer curves (Jahnke et al., 10 Mar 2025). The $\mathcal{P}_e$14 framework identifies cost modeling as a major open issue (Yan et al., 2016). EdgePoint2 does not introduce a literal metric called EEI, and its numeric “EEI-like” examples are explicitly illustrative rather than canonical (Yao et al., 24 Apr 2025). Edge proximity is less informative in very small or very dense graphs (Banerjee et al., 2014).

Taken together, these works support a precise general interpretation: an EEI is a normalized measure of useful edge-side function relative to the dominant limiting resource of the system under study. The exact form depends on whether the edge object is a descriptor network, an edge–cloud deployment, a radio-access infrastructure with caching and X-Haul, or an individual edge in a graph.