XAI-on-RAN: Explainable, AI-native, and GPU-Accelerated RAN Towards 6G

Falko Dressler; Osman Tugay Basaran

REVIEW 4 major objections 4 minor 18 references

A GPU-accelerated O-RAN can run real-time explainable AI within the 10 ms control loop using a hybrid attention-and-integrated-gradients xApp, the paper argues.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 11:17 UTC pith:L6PLZXRT

load-bearing objection Real XAI xApp on a GPU RAN, but the headline fidelity win over SHAP looks like it may be an under-sampling artifact. the 4 major comments →

arxiv 2511.17514 v1 pith:L6PLZXRT submitted 2025-10-07 cs.NI cs.AIcs.ITmath.IT

XAI-on-RAN: Explainable, AI-native, and GPU-Accelerated RAN Towards 6G

Osman Tugay Basaran , Falko Dressler This is my paper

classification cs.NI cs.AIcs.ITmath.IT

keywords explainable AIO-RANRAN intelligent controllerintegrated gradientsattention mechanismGPU acceleration6Gnetwork management

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that explainable AI can be made fast enough to guide every decision in a 6G open radio access network (O-RAN) without blowing the control-loop budget. The authors build a testbed where a dedicated xApp on the near-real-time RAN intelligent controller generates explanations for a traffic predictor's inferences on the same GPU that handles the physical layer. They model the trade-off between explanation fidelity, latency, and GPU utilization, and benchmark three approaches: attention, integrated gradients, and SHAP. Their hybrid xAI-Native method, attention plus integrated gradients, delivers explanations in 8.1 ms total per inference cycle while achieving the highest local fidelity, suggesting that real-time transparency is feasible for mission-critical network control.

Core claim

On the paper's own terms, the central discovery is that combining intrinsic attention weights with post-hoc integrated-gradients attributions, run on an NVIDIA A100 GPU inside the O-RAN near-RT RIC, yields the best fidelity–latency balance among the tested methods. The hybrid method explains a traffic predictor's output with an average total latency of 8.1 ms, staying within the 10 ms near-real-time budget, while feature-wise local fidelity (R^2_loc) is roughly 92% higher than SHAP and 37% higher than attention. SHAP, with 16 samples, costs about 20.4 ms total, which the authors consider borderline for ultra-reliable low-latency loops; attention alone is fast (0.6 ms overhead) but has low an

What carries the argument

The load-bearing mechanism is the xAI-Native xApp itself, which couples an attention layer intrinsic to the LSTM traffic predictor with an integrated-gradients post-hoc explainer (k=5 steps), all co-located on the GPU that processes L1/L2. The paper defines a local-fidelity metric R^2_loc based on a linear surrogate (Eq. 1) and a simpler top-k fidelity score Φ, plus a latency decomposition T_total = T_inf + T_xai + T_comm. These two measurements — fidelity and latency — are what let the paper compare methods on a single cost/benefit axis.

Load-bearing premise

The central claim rests on the assumption that the local linear-surrogate fidelity R^2_loc (and the top-k fidelity score) actually measures the transparency and trust that operators and regulators need; if high R^2_loc does not correspond to explanations people can act on, the fidelity–latency ranking loses its meaning.

What would settle it

A controlled user study where operators are asked to detect injected model errors or biases using explanations from each method. If operators perform no better with IG attributions than with attention or SHAP, despite IG's higher R^2_loc, the paper's conclusion would fail. Alternatively, a replication on a different GPU or with a more complex model that pushes total latency above 10 ms would show the real-time claim is not general.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

If xAI-Native holds up, O-RAN operators can add explanation to every inference without leaving the near-real-time control loop, since the 8.1 ms total stays under typical 10-100 ms RIC budgets.
SHAP can be relegated to offline auditing and compliance, because its ~20 ms per explanation is too slow for live decisions but its theoretical guarantees remain useful outside the loop.
Attention-weight explanations, despite being nearly free, should not be trusted alone in RAN contexts; the paper shows they miss key features like BLER and SINR.
GPU-accelerated gradient-based post-hoc methods become the default for real-time XAI, with IG's overhead (2.8 ms for k=5 steps) a practical trade for the fidelity gain.
The modeling framework for the transparency–latency–GPU trade-off can be reused to evaluate other XAI methods or hardware configurations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Because the paper validates fidelity only via R^2_loc and a top-k error metric, not via human operators, its 'best fidelity' ranking may not carry over to actual trust or usability; an operator study would test whether IG explanations actually improve decision-making or error detection.
The 8.1 ms figure depends on a single A100 GPU with a small LSTM and periodic burst traffic; on shared or less capable GPUs, or with more features, the margin to 10 ms may vanish.
A natural extension would be to close the loop: feed the explanations back into the RIC policy (e.g., to suppress predictions the explainer doesn't support), turning XAI from a monitor into a controller.
The finding that attention alone is low-fidelity could generalize beyond this testbed and should caution against using attention-based interpretability claims for other RAN AI models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Real XAI xApp on a GPU RAN, but the headline fidelity win over SHAP looks like it may be an under-sampling artifact.

read the letter

Colleague,

You should know this paper is a genuine system-building effort. The authors ran a live XAI xApp inside an O-RAN near-RT RIC on an A100 GPU, with a real radio and a COTS UE on a robot. Their integrated-gradients path completes in 8.1 ms total, which is the kind of number the O-RAN community has been asking for. That part is real and worth citing as a pointer.

What’s new is C1: to my knowledge, the first real-time XAI xApp on a GPU-centric RAN. The latency breakdown and the feature-wise fidelity plots are useful raw material. The paper is honest about some things: it says SHAP is offline-only, and it acknowledges the debate about attention as explanation.

The soft spots are real, and they cluster in the comparison that carries the conclusions. The stress-test note is correct: Table 1’s +0.41 median advantage over SHAP is measured against a SHAP baseline with m=16 samples, and §4.3 admits SHAP ‘had some variance run-to-run’ and stabilized only with more samples. That means the headline dominance could be an artifact of comparing an exact IG attribution to an under-converged estimator. They need to report SHAP at a stabilized sample count, or frame m=16 as a fixed-latency budget and accept that they are comparing a budgeted SHAP, not SHAP per se.

Second, the ‘Attention + IG’ hybrid is never ablated. In Table 1 and the fidelity analysis, the hybrid reduces to IG, so the contribution of the attention component is untested. The label overstates the method.

Third, R²loc is a reasonable proxy but it is not validated against anything operators actually need — fairness audits, error detection, usability. That is a limitation, not a fatal flaw.

Minor things: SHAP latency is an offline estimate, not a live-loop measurement; the ‘mathematical framework’ in §3 is an additive latency model with symbols, not a framework; and there is no code or data posted, which is a problem for a system paper.

Who should read this: people working on XAI for O-RAN, real-time interpretability, and GPU-accelerated RAN. I would send it to peer review — the testbed is real and the question matters — but I would ask for a re-run of the SHAP baseline at a converged sample count, an ablation that isolates the attention component, and an explicit statement about the fidelity metric’s scope. As it stands, I would not cite the fidelity ranking; I would cite the system integration with a caveat.

Recommendation: engage, but treat F2 as provisional until the SHAP baseline is fixed.

Referee Report

4 major / 4 minor

Summary. The paper introduces XAI-on-RAN, a GPU-accelerated O-RAN testbed with an explainability xApp (xAI-Native) integrated into the near-RT RIC. The authors propose a hybrid Attention + Integrated Gradients (IG) approach, model the latency/fidelity trade-off with a simple additive decomposition (Section 3), and evaluate local fidelity (R²loc) and per-inference latency against SHAP and attention-only baselines (Section 4). They report that the hybrid achieves 8.1 ms total latency and the best fidelity–latency balance, concluding that SHAP remains suitable only for offline auditing and attention is low-fidelity.

Significance. If the central claims hold, this is a useful early demonstration of real-time XAI inside an O-RAN RIC: a working xApp on a real GPU-accelerated testbed, with measured latencies below 10 ms and a head-to-head comparison of three explanation methods. The contributions C1–C3 are concrete, and the paper is candid about SHAP's computational cost. However, the empirical support for the headline result is weakened by an under-converged SHAP baseline, the absence of an ablation isolating the 'hybrid' component, and the unvalidated use of R²loc as a stand-in for operator trust.

major comments (4)

[§4.1, Table 1, F1] The paper labels the proposed model 'Ours (Attention + IG, k=5)' and claims the hybrid combination is central (F1), but no experiment isolates the contribution of attention over IG alone. The fidelity analysis in §4.1 and Table 1 appears to compare IG attributions against SHAP and attention, not 'Attention+IG' against IG-only. Without an ablation, the paper does not establish that the hybrid architecture provides any fidelity gain beyond plain IG, which is load-bearing for F1 and F2.
[§4.3, §3.2, Table 1] The SHAP baseline is evaluated with m=16 samples for a 5-feature model, while exact Shapley values require 2^5=32 evaluations. The manuscript itself states in §4.3 that 'the SHAP estimates had some variance run-to-run. With more samples, they stabilized.' The reported median ΔR²loc = +0.41 over SHAP may therefore reflect comparison against an under-converged estimator rather than a genuine fidelity advantage. Please report SHAP at a converged sample count (e.g., exact Shapley for 5 features) or justify why m=16 is sufficient.
[§3.1, Eq. (1), F2] The central conclusion (F2) ranks methods by R²loc, a local linear-surrogate fidelity metric. The paper never validates that high R²loc corresponds to explanations that operators actually find transparent, usable, or useful for error/fairness auditing. Without such validation—e.g., a task-based operator study or a downstream audit metric—the statement that the xApp provides 'human-interpretable insights' and the 'best fidelity–latency balance' rests on an unproven premise.
[§4.2, Table 2] The latency comparison is partly asymmetric: the proposed Attention+IG row is measured in the live loop (8.1 ms total), whereas the SHAP row is an offline projection ('we did not run it in the live loop, but timed it offline ... estimated 15 ms extra'). This makes the SHAP-vs-ours latency gap less reliable than the presentation implies. Please report whether T_comm is included in the XAI rows and clearly separate measured from projected numbers.

minor comments (4)

[§4.1, Figure 2] The manuscript refers to 'feature-wise' and 'time-wise' fidelity but does not define the perturbation neighborhood, the sliding-window size W, or the exact computation of R²loc in Eq. (1). These details are needed for reproducibility.
[§4.3] The admission that SHAP estimates stabilize only with more samples is important; consider moving it earlier and making the convergence behavior a quantitative part of the comparison.
[Throughout] Typographical issues: 'fideliy' in §4.1, 'T(xal)' in Table 2 header, and inconsistent use of 'R²loc' vs 'R2loc'. Also, the claim that backward is faster than forward is model-specific and should be justified or softened.
[§2] The architecture description is clear, but a block diagram of the xAI-Native xApp data flow (message subscription, explanation generation, dashboard output) would help readers understand the live-loop operation.

Circularity Check

0 steps flagged

No circularity: fidelity and latency claims rest on measurements, not on inputs re-labeled as predictions; self-citations are infrastructural and non-load-bearing.

full rationale

The paper's derivation chain is empirical rather than definitional. The central comparison (Table 1, Figure 2) uses measured R²loc values computed from the same local linear surrogate definition (Eq. 1) applied to all three explanation methods; IG's higher R²loc is an empirical outcome, not an identity, because SHAP and attention attributions are inserted into the same surrogate and could in principle score higher. The latency results in Table 2 are direct measurements, with the IG overhead (2.8 ms) checked against the k·Tinf-back estimate rather than used to define that estimate. Self-citations [9] and [10] provide prior infrastructure and motivation, but they are not invoked as evidence for F1/F2 and no argument reduces to them. The one admitted weakness is in Sec. 4.3: 'because we used only 16 samples, the SHAP estimates had some variance run-to-run. With more samples, they stabilized' — this is a benchmark-convergence limitation that affects the fairness of the IG-vs-SHAP fidelity comparison, but it is not a circular step: no fitted parameter is renamed as a prediction, and no equation reduces to its own input. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper does not introduce new physical or mathematical entities. Its framework rests on chosen parameters (k=5 IG steps, m=16 SHAP samples, α≈0.1, γ≈0.5, 80% coverage) and on domain assumptions about fidelity measurement and latency decomposition. None of these is independently validated, but none is circular in the sense of assuming the target result.

free parameters (5)

IG integration steps k = 5
Chosen for the proposed xAI-Native design (Tables 1 and 2); IG overhead scales with k, so the latency result is conditional on this choice.
SHAP sample count m = 16
Chosen as the maximum manageable number of forward passes; the SHAP latency estimate (~15 ms) is m·Tinf and would grow beyond the real-time budget if m increased.
Attention overhead fraction alpha_attn = ≈0.1
Postulated in Eq. (4) as 'on the order of 10%' of inference time; not measured in an ablation.
Backward/forward time ratio gamma = ≈0.5
Used to reconcile measured IG overhead (2.8 ms ≈ 5.2 × 0.5) in §4.2; a post-hoc characterization of this specific model.
Attribution coverage threshold for top-k = 80%
Chosen so the top features 'account for say 80% of total attribution weight' (§3.1); the fidelity score depends on this choice.

axioms (4)

domain assumption Local linear surrogate R² is a valid measure of explanation fidelity
Eq. (1) defines R²loc via a linear surrogate; used throughout §4.1 to rank explanations, but never validated against human/operator interpretability or task success.
domain assumption Attention weights can be treated as feature-importance explanations
§4.3 acknowledges the debate but adopts attention as an explanation baseline for the LSTM; if attention is not a valid explanation, the Attention-vs-IG comparison is weakened.
domain assumption Latency components decompose additively and Tcomm is negligible on the co-located GPU/RIC
§3.2 sets Ttotal = Tinf + Txai + Tcomm and assumes co-location makes Tcomm negligible; this ignores queuing and synchronization effects.
domain assumption GPU utilization percentage is a meaningful headroom/load metric
Table 2 uses utilization to claim the GPU 'still wasn't fully saturated'; utilization depends on measurement window and batching choices not described.

pith-pipeline@v1.3.0-alltime-deepseek · 7352 in / 15212 out tokens · 128560 ms · 2026-08-04T11:17:41.731897+00:00 · methodology

0 comments

read the original abstract

Artificial intelligence (AI)-native radio access networks (RANs) will serve vertical industries with stringent requirements: smart grids, autonomous vehicles, remote healthcare, industrial automation, etc. To achieve these requirements, modern 5G/6G design increasingly leverage AI for network optimization, but the opacity of AI decisions poses risks in mission-critical domains. These use cases are often delivered via non-public networks (NPNs) or dedicated network slices, where reliability and safety are vital. In this paper, we motivate the need for transparent and trustworthy AI in high-stakes communications (e.g., healthcare, industrial automation, and robotics) by drawing on 3rd generation partnership project (3GPP)'s vision for non-public networks. We design a mathematical framework to model the trade-offs between transparency (explanation fidelity and fairness), latency, and graphics processing unit (GPU) utilization in deploying explainable AI (XAI) models. Empirical evaluations demonstrate that our proposed hybrid XAI model xAI-Native, consistently surpasses conventional baseline models in performance.

Figures

Figures reproduced from arXiv: 2511.17514 by Falko Dressler, Osman Tugay Basaran.

**Figure 2.** Figure 2: Comparative evaluation with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 1 linked inside Pith

[1]

Goodfellow, Y

I. Goodfellow, Y . Bengio, and A. Courville, Deep learning. MIT Press, 2016

2016
[2]

Empowering the 6G Cellular Architecture with Open RAN,

M. Polese, M. Dohler, F. Dressler, M. Erol-Kantarci, R. Jana, R. Knopp, and T. Melodia, “Empowering the 6G Cellular Architecture with Open RAN,” IEEE Journal on Selected Areas in Communications, vol. 42, no. 2, pp. 245–262, Feb. 2024

2024
[3]

Understanding O-RAN: Archi- tecture, Interfaces, Algorithms, Security, and Research Challenges,

M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Understanding O-RAN: Archi- tecture, Interfaces, Algorithms, Security, and Research Challenges,”IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023

2023
[4]

Explainable Artificial Intelligence for 6G: Improving Trust between Human and Machine,

W. Guo, “Explainable Artificial Intelligence for 6G: Improving Trust between Human and Machine,” IEEE Communications Magazine, vol. 58, no. 6, pp. 39–45, Jun. 2020

2020
[5]

A Survey on Explainable AI for 6G O-RAN: Architecture, Use Cases, Challenges and Research Directions,

B. Brik, H. Chergui, L. Zanzi, F. Devoti, A. Ksentini, M. S. Siddiqui, X. Costa-Pérez, and C. Verikoukis, “A Survey on Explainable AI for 6G O-RAN: Architecture, Use Cases, Challenges and Research Directions,” arXiv, cs.NI 2307.00319, Jul. 2023

Pith/arXiv arXiv 2023
[6]

EXPLORA: AI/ML EXPLainability for the Open RAN,

C. Fiandrino, L. Bonati, S. D’Oro, M. Polese, T. Melodia, and J. Widmer, “EXPLORA: AI/ML EXPLainability for the Open RAN,” Proceedings of the ACM on Networking, vol. 1, pp. 1–26, Nov. 2023

2023
[7]

R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction. Cambridge, MA: MIT Press, 1998, p. 322

1998
[8]

An Introduction to Deep Reinforcement Learning,

V . François-Lavet, P. Henderson, R. Islam, M. G. Bellemare, and J. Pineau, “An Introduction to Deep Reinforcement Learning,” Foundations and Trends® in Machine Learning, vol. 11, no. 3–4, pp. 219–354, 2018

2018
[9]

XAInomaly: Explainable and Interpretable Deep Contractive Autoencoder for O-RAN Traffic Anomaly Detection,

O. T. Basaran and F. Dressler, “XAInomaly: Explainable and Interpretable Deep Contractive Autoencoder for O-RAN Traffic Anomaly Detection,” Elsevier Computer Networks, vol. 261, p. 111 145, Apr. 2025

2025
[10]

Next-Gen AI-on-RAN: AI-native, Interoperable, and GPU-Accelerated Testbed Towards 6G Open-RAN,

O. T. Basaran, H. Zafar, M. Kasparick, F. Dressler, and S. Sta´nczak, “Next-Gen AI-on-RAN: AI-native, Interoperable, and GPU-Accelerated Testbed Towards 6G Open-RAN,” inIEEE International Conference on Communications (ICC 2025), Montréal, Canada: IEEE, Jun. 2025, pp. 5056–5061

2025
[11]

OSC Near Realtime RIC

O.-R. SC. “OSC Near Realtime RIC.” (2024), [Online]. Available: https://wiki.o-ran- sc.org/display/RICP/. Accessed: 04.06.2025

2024
[12]

NVIDIA Aerial SDK

NVIDIA. “NVIDIA Aerial SDK.” (2024), [Online]. Available: https : / / developer . nvidia.com/aerial-%20sdk. Accessed: 04.06.2025

2024
[13]

CUDA Toolkit

NVIDIA. “CUDA Toolkit.” (2024), [Online]. Available: https : / / developer . nvidia . com/cuda-toolkit. Accessed: 04.06.2025

2024
[14]

Foxconn RPQN

Foxconn. “Foxconn RPQN.” (2024), [Online]. Available: https : / / fcc . report / FCC - ID/2AQ68RPQN7801/5573870.pdf. Accessed: 04.06.2025

2024
[15]

OpenAirInterface: Democratizing innovation in the 5G Era,

F. Kaltenberger, A. P. Silva, A. Gosain, L. Wang, and T. -T. Nguyen, “OpenAirInterface: Democratizing innovation in the 5G Era,” Elsevier Computer Networks, vol. 176, p. 107 284, Jul. 2020

2020
[16]

Grafana Monitoring

Grafana. “Grafana Monitoring.” (2024), [Online]. Available:https://grafana.com/docs/. (accessed: 10.08.2024). 7

2024
[17]

“Why Should I Trust You?

M. T. Ribeiro, S. Singh, and C. Guestrin, ““Why Should I Trust You?”: Explaining the Predictions of Any Classifier,” in22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA: ACM, Aug. 2016, pp. 1135–1144

2016
[18]

A Unified Approach to Interpreting Model Predictions,

S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” in 31st International Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA: Curran Associates Inc., Dec. 2017, pp. 4768–4777. 8

2017

[1] [1]

Goodfellow, Y

I. Goodfellow, Y . Bengio, and A. Courville, Deep learning. MIT Press, 2016

2016

[2] [2]

Empowering the 6G Cellular Architecture with Open RAN,

M. Polese, M. Dohler, F. Dressler, M. Erol-Kantarci, R. Jana, R. Knopp, and T. Melodia, “Empowering the 6G Cellular Architecture with Open RAN,” IEEE Journal on Selected Areas in Communications, vol. 42, no. 2, pp. 245–262, Feb. 2024

2024

[3] [3]

Understanding O-RAN: Archi- tecture, Interfaces, Algorithms, Security, and Research Challenges,

M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Understanding O-RAN: Archi- tecture, Interfaces, Algorithms, Security, and Research Challenges,”IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023

2023

[4] [4]

Explainable Artificial Intelligence for 6G: Improving Trust between Human and Machine,

W. Guo, “Explainable Artificial Intelligence for 6G: Improving Trust between Human and Machine,” IEEE Communications Magazine, vol. 58, no. 6, pp. 39–45, Jun. 2020

2020

[5] [5]

A Survey on Explainable AI for 6G O-RAN: Architecture, Use Cases, Challenges and Research Directions,

B. Brik, H. Chergui, L. Zanzi, F. Devoti, A. Ksentini, M. S. Siddiqui, X. Costa-Pérez, and C. Verikoukis, “A Survey on Explainable AI for 6G O-RAN: Architecture, Use Cases, Challenges and Research Directions,” arXiv, cs.NI 2307.00319, Jul. 2023

Pith/arXiv arXiv 2023

[6] [6]

EXPLORA: AI/ML EXPLainability for the Open RAN,

C. Fiandrino, L. Bonati, S. D’Oro, M. Polese, T. Melodia, and J. Widmer, “EXPLORA: AI/ML EXPLainability for the Open RAN,” Proceedings of the ACM on Networking, vol. 1, pp. 1–26, Nov. 2023

2023

[7] [7]

R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction. Cambridge, MA: MIT Press, 1998, p. 322

1998

[8] [8]

An Introduction to Deep Reinforcement Learning,

V . François-Lavet, P. Henderson, R. Islam, M. G. Bellemare, and J. Pineau, “An Introduction to Deep Reinforcement Learning,” Foundations and Trends® in Machine Learning, vol. 11, no. 3–4, pp. 219–354, 2018

2018

[9] [9]

XAInomaly: Explainable and Interpretable Deep Contractive Autoencoder for O-RAN Traffic Anomaly Detection,

O. T. Basaran and F. Dressler, “XAInomaly: Explainable and Interpretable Deep Contractive Autoencoder for O-RAN Traffic Anomaly Detection,” Elsevier Computer Networks, vol. 261, p. 111 145, Apr. 2025

2025

[10] [10]

Next-Gen AI-on-RAN: AI-native, Interoperable, and GPU-Accelerated Testbed Towards 6G Open-RAN,

O. T. Basaran, H. Zafar, M. Kasparick, F. Dressler, and S. Sta´nczak, “Next-Gen AI-on-RAN: AI-native, Interoperable, and GPU-Accelerated Testbed Towards 6G Open-RAN,” inIEEE International Conference on Communications (ICC 2025), Montréal, Canada: IEEE, Jun. 2025, pp. 5056–5061

2025

[11] [11]

OSC Near Realtime RIC

O.-R. SC. “OSC Near Realtime RIC.” (2024), [Online]. Available: https://wiki.o-ran- sc.org/display/RICP/. Accessed: 04.06.2025

2024

[12] [12]

NVIDIA Aerial SDK

NVIDIA. “NVIDIA Aerial SDK.” (2024), [Online]. Available: https : / / developer . nvidia.com/aerial-%20sdk. Accessed: 04.06.2025

2024

[13] [13]

CUDA Toolkit

NVIDIA. “CUDA Toolkit.” (2024), [Online]. Available: https : / / developer . nvidia . com/cuda-toolkit. Accessed: 04.06.2025

2024

[14] [14]

Foxconn RPQN

Foxconn. “Foxconn RPQN.” (2024), [Online]. Available: https : / / fcc . report / FCC - ID/2AQ68RPQN7801/5573870.pdf. Accessed: 04.06.2025

2024

[15] [15]

OpenAirInterface: Democratizing innovation in the 5G Era,

F. Kaltenberger, A. P. Silva, A. Gosain, L. Wang, and T. -T. Nguyen, “OpenAirInterface: Democratizing innovation in the 5G Era,” Elsevier Computer Networks, vol. 176, p. 107 284, Jul. 2020

2020

[16] [16]

Grafana Monitoring

Grafana. “Grafana Monitoring.” (2024), [Online]. Available:https://grafana.com/docs/. (accessed: 10.08.2024). 7

2024

[17] [17]

“Why Should I Trust You?

M. T. Ribeiro, S. Singh, and C. Guestrin, ““Why Should I Trust You?”: Explaining the Predictions of Any Classifier,” in22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA: ACM, Aug. 2016, pp. 1135–1144

2016

[18] [18]

A Unified Approach to Interpreting Model Predictions,

S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” in 31st International Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA: Curran Associates Inc., Dec. 2017, pp. 4768–4777. 8

2017