The Price of Meaning: Quantifying Semantic Communication Overheads in Practice Xinyi Lin, Peizheng Li, Adnan Aijaz
arXiv:2607.26764v1 [cs.NI] 29 Jul 2026
Bristol Research and Innovation Laboratory, Toshiba Europe Ltd., U.K. Email: {xinyi.lin, peizheng.li, adnan.aijaz}@toshiba-bril.com
Abstract—Semantic communication (SemCom) promises to reduce transmitted payloads by conveying task-relevant meaning instead of raw bits. However, practical SemCom also incurs semantic metadata, control signaling, feedback, model or knowledge-base synchronization, and neural computation costs, which may offset semantic compression gains. This paper develops an overhead-aware analytical framework for quantifying the spectral-resource and energy costs of SemCom under equal task utility. The framework covers point-to-point transmission, user equipment (UE)-to-next-generation NodeB (gNB) uplink, and UE-to-UE communication under a single gNB, and derives closed-form break-even conditions with respect to payload size, semantic compression factor, model reuse, protocol overhead, and computation energy. Simulation results show that SemCom becomes spectrally beneficial only for sufficiently large payloads, while energy gains require larger payloads due to processing and synchronization overheads. The results also show that multiuser downlink is particularly favorable, as shared semantic overheads can be amortized across multiple UEs. These findings provide design guidance for realistic SemCom evaluation and standardization-oriented deployment. Index Terms—Semantic communication, 6G, 5G NR, overhead analysis, energy efficiency, spectral efficiency, model synchronization.
overheads, including scheduling, uplink grants, physical uplink control channel (PUCCH), physical downlink control channel (PDCCH), demodulation reference signals (DMRS), hybrid automatic repeat request (HARQ), and radio resource control (RRC) procedures [5]–[7]. Thus, reducing the application payload does not automatically imply reducing the total radio or energy cost. This paper develops a compact analytical model for overhead-aware SemCom evaluation. Rather than asking whether semantic representations are shorter than conventional payloads, we ask whether SemCom reduces the total cost needed to achieve a target utility. We consider both communication-resource and energy cost, and we instantiate the model in three deployment scenarios: point-to-point communication, UE-to-gNB uplink communication, and UE-to-UE communication within a single gNB coverage area. The resulting expressions expose the conditions under which SemCom is beneficial and the regimes where overhead dominates.
I. I NTRODUCTION EMANTIC communication (SemCom) revisits the classical communication objective by emphasizing the transfer of meaning, task utility, or receiver action rather than bitexact reconstruction. Early information theory deliberately separated the technical transmission problem from semantic and effectiveness problems [1], [2]; modern SemCom reopens this separation by using learned representations, knowledge bases, and task-oriented metrics. Neural SemCom systems such as DeepSC have shown that semantic representations can be robust at low signal-to-noise ratio (SNR) and may reduce the amount of payload that must be transmitted [3]. Surveys and tutorials further highlight SemCom as a candidate enabler for 6G intelligence, sensing, and edge-native applications [4]. Despite these advances, a key practical question remains unresolved: what is the price of meaning? A semantic payload is not transmitted in isolation. Real systems must identify the semantic task, model version, embedding format, context, confidence level, and reliability target. They may also require semantic feedback, channel feedback, retransmission triggers, and model or knowledge-base synchronization. In cellular deployments, these semantic-specific costs coexist with unavoidable 5G New Radio (NR) control and reference-signal
Existing SemCom studies can be grouped into four categories. First, foundational and tutorial works define the SemCom vision, semantic metrics, knowledge support, and implementation choices [4]. Second, neural end-to-end systems, including text, speech, and image SemCom, demonstrate robustness and payload reduction relative to conventional bit-oriented baselines [3]. Third, resource-allocation studies introduce semantic spectral efficiency and optimize channel or semantic-symbol allocation [8]. Fourth, recent energy-aware works account for the cost of semantic extraction, inference, or communication–computation tradeoffs. For example, ratesplitting-based SemCom minimizes communication and computation energy [9], green transformer selection benchmarks semantic loss against CPU/GPU energy [10], and probabilistic SemCom over space-air-ground integrated networks models the tradeoff between semantic compression and computation energy [11]. Feedback-aware SemCom has also been studied for reliability, where adaptive channel feedback is allocated based on predicted semantic distortion [12]. These works establish that SemCom can improve task performance, spectral efficiency, or energy efficiency under suitable assumptions. However, most evaluations focus on semantic payloads, semantic symbols, transmit energy, or
S
II. R ELATED W ORK AND C ONTRIBUTIONS
inference energy. They usually do not jointly quantify the practical overhead stack required for deployment: protocol control signaling, semantic metadata, reliability feedback, model or knowledge-base synchronization, and UE-side computation. This gap is especially important for short packets, where fixed control and synchronization overhead can exceed the semantic payload itself. This paper makes the following contributions: • We develop an overhead-aware SemCom cost framework that jointly accounts for semantic payload compression, metadata, control signaling, feedback, reference signals, model/knowledge synchronization, and endpoint computation. • We instantiate the framework for practical wireless deployments, including point-to-point links, NR UE– gNB uplink radio interface, gNB-to-multi-UE downlink, sidelink UE-to-UE communication, and network-routed UE-to-UE communication under one gNB. • We derive closed-form spectral-resource and energy break-even conditions that expose the roles of payload size, semantic compression factor, model reuse interval, protocol overhead, and neural computation energy. • We provide numerical results showing when SemCom is beneficial in practice: spectral gains appear after fixed overheads are amortized, energy gains require stricter conditions, multi-user downlink offers strong overhead sharing, and large model reuse with strong compression is critical. III. S YSTEM M ODEL A. Utility-Constrained Communication Consider a source message associated with a task utility target U0 . A conventional transmitter sends a source-coded payload of length L bits. A semantic transmitter maps the source to a task-relevant representation using encoder fθ and decoder gϕ with shared context or knowledge base K, where θ and ϕ are the encoder and decoder model parameters. For a target utility U0 , define the semantic compression factor 0 < ρ(U0 ) ≤ 1,
(1)
B. Generic Overhead Taxonomy The conventional total information-equivalent burden is (s)
(s)
(s)
tot BC = L + BC,ctrl + BC,fb + BC,hdr , (s)
(s)
(4)
(s)
where BC,ctrl , BC,fb and BC,hdr denote the control-channel, feedback, and protocol header burden, respectively. For SemCom, the burden is (s)
(s)
(s)
(s)
BStot = ρL + Bmeta + BS,ctrl + BS,fb +
Bsync , N
(5)
(s)
where Bmeta includes model identifier, task identifier, embedding format, context index, confidence information, semantic quality-of-service (QoS) fields, and tokenizer or codebook (s) version. The term Bsync is the traffic required to distribute or update the semantic model, codebook, probability graph, or knowledge base, and N is the number of payloads over which this cost is amortized. Thus, small N represents dynamic tasks or frequent model updates, while large N represents stable shared context. C. Spectral-Resource Cost Let Hs denote the set of wireless hops in scenario s. The total spectral-resource cost is modeled as pay meta X Bx,ℓ + Bx,ℓ (s) (s) + Ax,ctrl + Ax,fb A(s) = x ηx,ℓ ℓ∈Hs
(s)
Ax,sync , (6) N pay where ηx,ℓ is the payload spectral efficiency on hop ℓ, Bx,ℓ meta and Bx,ℓ are the payload and metadata bits transmitted over (s) (s) hop ℓ, Ax,ctrl is the control-channel resource cost, Ax,fb is (s) the feedback resource cost, Ax,rs accounts for pilots, DMRS, (s) or other reference signals, and Ax,sync is the spectral cost of model or knowledge-base synchronization. Eq. (6) intentionally separates payload compression from fixed and semi-fixed protocol costs. + A(s) x,rs +
D. Energy Cost The total energy cost is
so that the semantic payload is BS,pay = ρ(U0 )L,
(s)
(2)
whereas the conventional payload is BC,pay = L. The parameter ρ(U0 ) captures the fact that stricter utility, distortion, or task-accuracy requirements reduce semantic compressibility. Let s ∈ {P2P, Uu, SL} index the deployment scenario, corresponding to point-to-point (P2P), NR Uu, and sidelink (SL) communication, respectively. Let x ∈ {C, S} index the communication mode, where C and S denote conventional and semantic communication. The evaluation target is Ux Ux (s) (s) ξeff,x = (s) , (3) ηeff,x = (s) , Ax Ex (s)
where Ux is the achieved utility, Ax is the total spectral(s) resource cost, and Ex is the total energy cost. A fair comparison is made at equal utility, i.e., US ≥ U0 and UC ≥ U0 .
(s)
(s)
(s)
(s) Ex(s) = Ex,air + Ex,ctrl + Ex,fb + Ex,cmp + (s)
(s)
(s)
(s)
Ex,sync , N
(7)
(s)
where Ex,air , Ex,ctrl , Ex,fb , Ex,cmp , and Ex,sync denote the air-interface, control signaling, feedback, computation, and synchronization energy, respectively, with the synchronization cost amortized over N payloads. The air-interface energy is X (s) tx rx Pℓtx Tx,ℓ + Pℓrx Tx,ℓ , (8) Ex,air = ℓ∈Hs
where tx rx Tx,ℓ ≈ Tx,ℓ =
pay meta Bx,ℓ + Bx,ℓ
. (9) Wℓ ηx,ℓ In the air-interface energy model, Pℓtx and Pℓrx are the transmit tx rx and receive power on hop ℓ, Tx,ℓ and Tx,ℓ are the correspondpay meta ing transmission and reception duration, Bx,ℓ and Bx,ℓ
are the payload and metadata bits, Wℓ is the transmission bandwidth, and ηx,ℓ is the spectral efficiency. For neural SemCom, computation energy is (s)
ES,cmp = eMAC (Cenc + Cdec ) + emem Macc + Epost , (10) where Cenc and Cdec are encoder and decoder operation counts, eMAC is the energy per multiply–accumulate operation, Macc is the memory-access count, emem is the energy consumed per memory-access, and Epost covers task-specific post-processing. (s) In the numerical evaluation, Ex,cmp denotes endpoint computation needed to meet U0 . For conventional communication it includes source coding, decoding, and post-processing; for SemCom it also includes semantic encoding/decoding, inference, memory movement, and semantic decision processing. This separation avoids counting transmit-bit reduction as an energy gain unless computation and synchronization energy are also amortized, especially for battery-limited UEs.
where AC,hdr is the conventional header cost, AC,ack and AS,ack are acknowledgement (ACK) feedback, Ameta is semantic metadata, Aalign is semantic alignment signaling, Ars is reference-signal overhead, and Amodel is model synchronization overhead. Similarly, the energy costs are P2P EC =eC L + EC,hdr + EC,ack + EC,rs + EC,codec , (13)
ESP2P =eS ρL + Emeta + ES,ack + Ealign + ES,rs Emodel + Eenc + Edec + . (14) N Thus, point-to-point SemCom is spectrally beneficial only if L ρL Amodel + ∆AP2P < , oh + ηS N ηC
(15)
with ∆AP2P oh =Ameta + AS,ack + Aalign + AS,rs − AC,hdr − AC,ack − AC,rs .
(16)
E. Deployment Scenarios
B. Scenario B: UE-to-gNB
We consider three scenarios. 1) Scenario A: Point-to-Point: A transmitter communicates directly with a receiver over one hop, i.e., HP2P = {1}, which isolates semantic overhead from cellular scheduling overhead. The semantic-specific costs are metadata, semantic feedback, alignment messages, and amortized model synchronization. 2) Scenario B: UE-to-gNB over Uu: A UE transmits to a serving gNB through the NR Uu uplink. In addition to payload transmission on the physical uplink shared channel (PUSCH), the UE pays for scheduling requests, grants, buffer status reporting, PDCCH monitoring, PUCCH feedback, DMRS, HARQ, and higher-layer headers. SemCom further adds task/model metadata, semantic-quality reports, and UEside encoding energy. 3) Scenario C: UE-to-UE within a Single gNB: Two UEs communicate under one gNB. The primary case considered is gNB-controlled NR sidelink, where the useful payload is carried on the physical sidelink shared channel (PSSCH), sidelink control information is conveyed through PSCCH/SCI, and sidelink feedback may be conveyed through PSFCH. The gNB may allocate resources and distribute or validate semantic model state. A network-routed alternative, UE1 →gNB→UE2 , can be recovered by treating the path as two Uu hops plus relay processing.
1) UE-gNB uplink: For the conventional uplink from UE to gNB, the spectral cost is expressed as
IV. OVERHEAD -AWARE A NALYTICAL M ODEL A. Scenario A: Point-to-Point For conventional point-to-point communication, the spectral resource cost can be interpreted as L + AC,hdr + AC,ack + AC,rs . (11) AP2P = C ηC For SemCom, ρL Amodel +Ameta +AS,ack +Aalign +AS,rs + , (12) AP2P = S ηS N
AUu C =
L
+ ASR + Agrant + ABSR,C ηPUSCH,C + ADMRS + AHARQ,C + APDCP/RLC .
(17)
For semantic uplink, AUu S =
ρL ηPUSCH,S
+ ASR + Agrant + ABSR,S + ADMRS
Amodel,Uu . (18) N Here, PUSCH is the physical uplink shared channel, SR is the scheduling request, BSR is the buffer status report, PDCP/RLC denotes packet data convergence protocol/radio link control overhead, and Amodel,Uu is the Uu model-synchronization resource cost. The semantic metadata term is decomposed as + AHARQ,S + Asem-meta + Asem-fb +
Asem-meta =AtaskID + AmodelID + Aversion + Acodebook + AQoS + Aconfidence .
(19)
The UE-side energy comparison is Uu EC,UE =EPUSCH,C + EPUCCH,C + EPDCCH-mon
+ EC,codec ,
(20)
Uu ES,UE =EPUSCH,S + EPUCCH,S + EPDCCH-mon Emodel,Uu + Eenc + Esem-meta + . (21) N Here, PUCCH is the physical uplink control channel, PDCCHmon denotes physical downlink control channel monitoring, and Eenc and EC,codec denote semantic encoder and conventional codec energy, respectively. The common scheduling terms cancel only if SemCom uses the same scheduling mode and numerology as the conventional baseline. Otherwise, Uu ∆AUu ctrl and ∆Ectrl must be retained explicitly.
2) gNB-to-multi-UE downlink: For the conventional downlink, denoted by superscript gd, the gNB transmits an independent unicast stream to each UE. Let K denote the number of served UEs. The total spectral-resource consumption is L + ADCI + ADMRS Agd C =K ηPDSCH,C ! + AHARQ,C + APDCP/RLC ,
(22)
where PDSCH is the physical downlink shared channel and ADCI denotes the spectral cost for downlink control information (DCI). For semantic downlink, a common semantic representation is transmitted once and reused by multiple UEs with similar semantic objectives. The corresponding spectralresource consumption becomes ρL Agd +ADCI +ADMRS +AHARQ,S +Asem-meta S = ηPDSCH,S Amodel,gd + K AUE-specific , (23) + Asem-fb + N where AUE-specific denotes the per-UE signaling that cannot be shared, such as UE-specific feedback, control signaling, or reliability-related procedures. The corresponding gNB-side energy consumption is gd EC,gNB = K (EPDSCH,C + EC,ctrl + EC,codec ) ,
(24)
gd =EPDSCH,S + ES,ctrl + Eenc ES,gNB
SL C C EC =EUE1,tx + EUE2,rx + EPSCCH,C
+ EPSFCH,C + EUu-sched,C + EC,codec ,
where PSSCH is the physical sidelink shared channel, PSCCH carries sidelink control, SCI carries sidelink control information, PSFCH carries sidelink feedback, SL-DMRS denotes sidelink demodulation reference signals, and AUu-sched captures gNB scheduling assistance. For semantic sidelink, ρL + APSCCH,S + ASCI,S + APSFCH,S ASL S = ηPSSCH,S + ASL-DMRS + AUu-sched,S + ASL sem-meta A model,SL + ASL . (27) sem-fb + N Here, ASL sem-meta =AtaskID + AmodelID + Aversion + AfeatureFormat (28)
(29)
S S ESSL = EUE1,tx + EUE2,rx + EPSCCH,S + EPSFCH,S Emodel,SL + EUu-sched,S + Eenc,UE1 + Edec,UE2 + . (30) N 2) UE-to-gNB-to-UE routed: For the network-routed alternative UE1 →gNB→UE2 , the same notation gives Uu Arouted = AUu x x,UL + Ax,DL + Ax,relay ,
(31)
where UL, DL, and relay denote uplink, downlink, and gNB relay-processing costs. An analogous expression applies for energy. This alternative increases both control and receive energy at the gNB, but may reduce direct sidelink coordination requirements. V. B REAK -E VEN A NALYSIS A. Spectral Break-Even Define the scenario-dependent overhead difference (s)
(s)
(s)
(s)
∆Aoh =AS,meta + AS,fb − AC,fb (s)
(s)
(s)
(s)
+ AS,ctrl − AC,ctrl + AS,rs − AC,rs .
(32)
Assuming a single effective payload spectral efficiency for each scenario s, SemCom is spectrally beneficial when (s)
Emodel,gd + Esem-meta + K EUE-specific + . (25) N Unlike the conventional downlink, where the payload transmission scales linearly with the number of users, semantic communication transmits a shared semantic representation only once. Therefore, only the UE-specific signaling and feedback overheads grow approximately linearly with K, while the semantic payload and model synchronization costs are amortized over multiple users. C. Scenario C: UE-to-UE under One gNB 1) UE-to-UE sidelink: For gNB-controlled sidelink, L ASL + APSCCH,C + ASCI,C C = ηPSSCH,C + APSFCH,C + ASL-DMRS + AUu-sched,C , (26)
+ AsemanticQoS + AUEcontext .
The sidelink energy costs are
(s)
AS < AC .
(33)
Substituting (6) gives the payload-size condition (s) (s) ∆Aoh + AS,sync /N A,(s) L > Lmin = . (s) (s) 1/ηC − ρ/ηS
(34)
Equation (34) is meaningful only when 1 ρ > (s) . (35) (s) ηC ηS If this condition fails, the semantic payload is not sufficiently compact relative to its spectral efficiency, and no amount of payload scaling can compensate for positive overhead. B. Energy Break-Even (s)
(s)
Let eC and eS denote effective communication energy per payload bit for conventional and semantic transmission, respectively. Let (s)
(s)
(s)
(s)
∆Eoh =ES,meta + ES,fb − EC,fb (s)
(s)
(s)
(s)
+ ES,ctrl − EC,ctrl + ES,rs − EC,rs .
(36)
Then SemCom is energy beneficial if (s) (s) (s) (s) ∆Eoh + ES,cmp − EC,cmp + ES,sync /N E,(s) L > Lmin = . (s) (s) eC − ρeS
(37)
The denominator requires (s)
(s)
eC > ρeS .
(38)
Thus, even if SemCom reduces transmitted payload, it may fail the energy test when neural inference, memory movement, synchronization, or receive-side processing dominates.
Equation (39) is the central design rule of this paper. It states that SemCom is not inherently efficient; rather, it becomes efficient only after the semantic payload savings amortize practical overheads. The most important variables are the semantic compression factor ρ, synchronization reuse factor N , (s) neural compute energy ES,cmp , and scenario-specific control (s) overhead ∆Aoh . D. Design Implications The analytical expressions provide several protocol-level implications. First, short-packet SemCom is vulnerable to fixed metadata, feedback, synchronization, and scheduling costs, because L may be smaller than (39). Second, stable applications with long model or knowledge-base reuse intervals are more favorable, since synchronization terms scale as 1/N . Third, UE-centric SemCom must be evaluated using device-side energy, because inference and memory access can dominate transmit-bit savings. For future 3GPP evolution, e.g., Release 20 and beyond toward 6G, SemCom should be standardized as an overheadaware protocol capability rather than only an application-layer compression method. This suggests lightweight signaling for semantic task IDs, model or knowledge-base IDs, versioning, representation formats, and semantic QoS targets. Semantic feedback should also be distinguished from conventional ACK/NACK, allowing receivers to report task-level utility, confidence, or semantic failure. Finally, model and knowledgebase synchronization should support cache validity, update periodicity, and multi-UE sharing, so that SemCom gains can be realized without ignoring metadata, feedback, computation, and synchronization costs.
Semantic / conventional spectral-resource cost
For equal utility U0 , SemCom is net-efficient in scenario s only if n o (s) A,(s) E,(s) L > Lmin = max Lmin , Lmin . (39)
TABLE I: Illustrative parameters for the numerical example. Parameter P2P Uu gd SL routed ηC 2.0 1.5 2.5 1.8 1.5 + 2.5 ηS 1.8 1.35 2.2 1.6 1.35 + 2.2 AC,oh 200 1200 500K 1600 2600 AS,oh 600 2500 4000 + 400K 3600 5200 6 6 6 6 Async 10 2 × 10 4 × 10 2.5 × 10 2.5 × 106 eC (nJ/bit) 0.002 0.010 0.006K 0.004 0.010+0.006 eS (nJ/bit) 0.002 0.010 0.006 0.004 0.010+0.006 EC,oh (nJ) 10 80 25K 120 180 ES,oh (nJ) 40 160 300 + 20K 240 340 EC,cmp (nJ) 20 20 10K 30 30 ES,cmp (nJ) 300 600 1000 + 400K 900 1000 5 5 5 5 Esync (nJ) 10 2 × 10 4 × 10 3 × 10 2.5 × 105
10
10
Break-even
0
P2P direct UE-to-gNB uplink gNB-to-10-UEs downlink UE-to-gNB-to-UE routed UE-to-UE sidelink
−1
10
2
10
3
10
4
10
5
10
6
Payload size, L [bits]
(a) Spectral cost ratio of SemCom over conventional communication. Semantic / conventional energy cost
C. Joint Efficiency Condition
10
1
10
0
Break-even
P2P direct UE-to-gNB uplink gNB-to-10-UEs downlink UE-to-gNB-to-UE routed UE-to-UE sidelink
10
2
10
3
10
4
10
5
10
6
Payload size, L [bits]
(b) Energy cost ratio of SemCom over conventional communication. Fig. 1: Resource cost ratio of SemCom over conventional communication.
VI. S IMULATION R ESULTS The numerical evaluation compares SemCom and conventional communication under equal utility. Unless otherwise stated, ρ = 0.3, K = 10, N = 1000, and L0 = 105 bits. The table columns P2P, Uu, gd, SL, and routed denote point-to-point, UE–gNB Uu, gNB downlink, sidelink, and network-routed UE-to-UE cases. The parameters ηC and ηS are conventional and semantic spectral efficiencies; AC,oh and AS,oh collect fixed protocol, metadata, feedback, and reference-signal overheads; Async is amortized synchronization overhead; eC and eS are energy per payload bit; and Eoh , Ecmp , and Esync are fixed overhead, computation, and synchronization energy costs. Table I summarizes the values. Fig. 1a and Fig. 1b compare the spectral and energy cost ratios between SemCom and conventional communication. Both ratios decrease as payload size increases, because fixed semantic overheads are amortized over more useful information. For short payloads, metadata, feedback, control, and scheduling costs can outweigh semantic compression gains, especially in
cellular scenarios. All considered scenarios eventually become spectrally beneficial, while the energy ratio remains above one over a wider range due to semantic encoding/decoding, metadata handling, and model synchronization energy. Fig. 2a and Fig. 2b show the amortization gain of multiuser downlink SemCom. As K increases, the spectral breakeven payload decreases rapidly because the common semantic representation and shared overheads are reused across more UEs, while the energy threshold decreases more gradually due to persistent encoder, synchronization, and UE processing costs. At L0 = 105 bits, the spectral ratio remains well below one and the energy ratio eventually crosses the break-even threshold, confirming that multi-user downlink is particularly favorable for SemCom. Fig. 3 further investigates the effect of the reuse factor N and the semantic compression ratio ρ on the break-even payload size. Fig. 3a illustrates the impact of the model reuse factor on the stricter break-even payload size. For all con-
Break-even payload size, L [bits]
Spectral break-even Energy break-even
10
5
10
4
10
3
0
5
10
15
20
25
30
Number of downlink UEs, K
(a) Multi-user downlink amortization gain under different numbers of UEs. Spectral ratio at 100 kbits Energy ratio at 100 kbits
Semantic / conventional cost ratio
3.5 3.0 2.5 2.0 1.5 Break-even
1.0 0.5 0.0 0
5
10
15
20
25
30
Number of downlink UEs, K
(b) Multi-user downlink cost ratio versus number of UEs.
Stricter break-even payload size [bits]
Fig. 2: Break-even payload size and semantic/conventional cost ratio under different numbers of UEs. 10
7
10
6
10
5
P2P direct UE-to-gNB uplink gNB-to-10-UEs downlink UE-to-gNB-to-UE routed UE-to-UE sidelink
10
1
10
2
10
3
10
4
10
5
10
6
Model / knowledge-base reuse factor, N
Stricter break-even payload size [bits]
(a) Reuse sensitivity of overhead-aware SemCom.
10
6
10
5
P2P direct UE-to-gNB uplink gNB-to-10-UEs downlink UE-to-gNB-to-UE routed UE-to-UE sidelink
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Semantic payload fraction, ρ
(b) Payload size sensitivity under different semantic compression factor ρ. Fig. 3: Break-even payload size versus reuse factor N and semantic compression factor ρ.
sidered scenarios, increasing the reuse factor N significantly reduces the break-even payload size, especially when the reuse factor is small. As N increases, the curves gradually converge, indicating that the synchronization cost becomes negligible and the remaining break-even payload size is dominated by transmission cost and fixed protocol overhead. Moreover, Fig. 3b shows the sensitivity of the break-even payload size to the semantic payload fraction ρ. As ρ increases, the transmission advantage of SemCom is weakened compared with conventional communication, leading to a larger break-even payload size. These results indicate that both model or knowledge-base reuse and semantic compression are key factors that determine the practicality of SemCom. VII. C ONCLUSION This paper quantified practical SemCom overhead using a unified spectral-resource and energy model across pointto-point, NR Uu, multi-user downlink, sidelink, and routed UE-to-UE deployments. The break-even conditions show that SemCom is beneficial only when semantic compression and model reuse amortize protocol, synchronization, and computation overheads. Numerical results confirm that short-payload SemCom is often inefficient, energy gains are stricter than spectral gains, and multi-user downlink, large model reuse, and strong compression are key enablers. ACKNOWLEDGMENT This work is supported by the 6G-GOALS project under the 6G SNS-JU Horizon program, n.101139232. R EFERENCES [1] C. E. Shannon, “A mathematical theory of communication,” The Bell System Technical Journal, vol. 27, no. 3, pp. 379–423, 1948. [2] C. E. Shannon and W. Weaver, The Mathematical Theory of Communication. Urbana, IL, USA: University of Illinois Press, 1949. [3] H. Xie, Z. Qin, G. Y. Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Transactions on Cognitive Communications and Networking, vol. 7, no. 3, pp. 746–757, 2021. [4] Z. Lu, R. Li, K. Lu, X. Chen, E. Hossain, Z.-F. Zhao, and H. Zhang, “Semantics-empowered communications: A tutorial-cum-survey,” IEEE Communications Surveys & Tutorials, vol. 26, no. 1, pp. 41–79, 2024. [5] 3GPP, “NR; multiplexing and channel coding,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 38.212, 2025, release 18. [6] ——, “NR; physical layer procedures for control,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 38.213, 2025, release 18. [7] ——, “NR; radio resource control (RRC); protocol specification,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 38.331, 2025, release 18. [8] X. Lin, P. Li, and A. Aijaz, “RL-Driven Semantic Compression Model Selection and Resource Allocation in Semantic Communication Systems,” in 2025 IEEE 36th International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), 2025, pp. 1–6. [9] Z. Yang, M. Chen, Z. Zhang, and C. Huang, “Energy efficient semantic communication over wireless networks with rate splitting,” arXiv preprint arXiv:2301.01987, 2023. [10] S. Mukherjee, C. C. Beard, and S. Song, “Transformers for green semantic communication: Less energy, more semantics,” arXiv preprint arXiv:2310.07592, 2023. [11] Z. Zhao, Z. Yang, M. Chen, Z. Zhang, W. Xu, and K. Huang, “Energyefficient probabilistic semantic communication over space-air-ground integrated networks,” arXiv preprint arXiv:2407.03776, 2024. [12] G. Zhang, Q. Hu, Y. Cai, and G. Yu, “SCAN: Semantic communication with adaptive channel feedback,” arXiv preprint arXiv:2306.15534, 2023.