Programmable and GPU-Accelerated Edge Inference for Real-Time ISAC on NVIDIA Aerial Testbed
Real-time ISAC on NVIDIA Aerial Testbed enables high-accuracy sensing for 6G with GPU-accelerated AI dApps.
This paper presents a groundbreaking framework that integrates GPU-accelerated Artificial Intelligence (AI) applications into the edge Radio Access Network (RAN) infrastructure for real-time Integrated Sensing and Communication (ISAC). Leveraging NVIDIA Aerial Testbed, the system processes PHY/MAC signals with minimal overhead (150 µs), supporting multiple inference engines and AI backends. A key demonstration, 'cuSense,' achieves 77 cm mean localization error for person tracking on a 5G NR deployment without dedicated sensing hardware, showcasing a practical pathway for AI-native RANs and 6G ISAC applications.
Key Executive Impact
Our framework delivers measurable performance advantages for next-generation RANs.
0
Framework Overhead
0
Localization Error
0
Predictions within 1m
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
150µs
Framework Overhead for Real-Time Inference on NVIDIA Aerial Testbed
GPU-Accelerated dApp Framework Data Path
cuPHY copies data GPU to CPU
→
cuBB notifies ADL data ready
→
ADL copies data CPU to SHM
→
E3 Agent sends pointers to E3 Manager
→
E3 Manager dispatches to engine
→
Engine sets inputs and runs inference
→
Engine returns results to E3 Manager
→
E3 Manager sends control to E3 Agent
→
E3 Agent receives & applies control
| Feature | Triton C API (TRT) | Triton gRPC (TRT) |
|---|---|---|
| Mean Latency |
|
|
| Serialization Overhead |
|
|
| Real-Time Suitability |
|
|
Case Study: cuSense: Indoor Person Localization
Challenge: Extracting accurate sensing estimates from communication signals (DMRS) in real-time with high-dimensional, noisy CSI data and static multipath components.
Solution: Developed cuSense, an ISAC dApp using real-time UL CSI estimates, static multipath removal, and a GPU-accelerated neural network for 2D position inference. Runs on NVIDIA Aerial Testbed without dedicated sensing hardware.
Result: Achieved a mean localization error of 77 cm (75% within 1 meter) in a 3GPP-compliant 5G NR deployment, meeting 3GPP sensing requirements.
77cm
Mean Localization Error of cuSense dApp
cuSense Uplink CSI Processing Pipeline
Input CSI Ht
→
Offline Background Avg. CSI Template HB
→
Temporal Averaging (Ht)
→
Background Filtering (Ht-HB)
→
Z-score Normalization
→
Neural Network Inference (2D Probability Map)
→
Kalman Filter Tracking
→
Raw Predictions
Case Study: GPU-Accelerated dApp Framework
Challenge: Enabling real-time, low-latency access to PHY/MAC data, co-locating dApps with gNB while ensuring loose coupling, providing GPU-native AI tooling, ensuring isolation and scalability, and aligning with O-RAN/AI-RAN standards.
Solution: Introduced a dApp framework on NVIDIA ATB 5G-NR stack with Real-time ADL using shared memory, an E3 Agent for communication, and a modular dApp container architecture supporting multiple AI backends.
Result: Achieved ~150 µs framework overhead and sub-millisecond E2E control-loop latency for real-time, high-accuracy ISAC dApps on GPU-accelerated RANs.
Calculate Your Potential ROI
Estimate the efficiency gains and cost savings your enterprise could achieve with GPU-accelerated AI.
Estimated Annual Savings
$0
Hours Reclaimed Annually
0
Your AI Implementation Roadmap
A typical journey to deploy GPU-accelerated AI for real-time applications.
Phase 1: Discovery & Strategy (2-4 weeks)
Initial assessment of existing infrastructure, data sources, and target use cases. Develop a tailored strategy aligned with business objectives, identifying key performance indicators and potential ROI. Includes stakeholder workshops and technology readiness evaluation.
Phase 2: Framework Integration & Pilot (6-10 weeks)
Integrate the GPU-accelerated dApp framework into your edge RAN. Develop and deploy a pilot ISAC or AI-native RAN application (e.g., cuSense localization) to validate the real-time data access and inference capabilities on a limited scale. Establish baseline performance metrics.
Phase 3: Model Development & Optimization (8-16 weeks)
Iterative development and training of AI/ML models using GPU-native tooling. Optimize models for low-latency inference, leveraging various backends (e.g., TRT, ONNX). Conduct extensive OTA evaluations and dataset refinement to ensure high accuracy and robust generalization.
Phase 4: Scalable Deployment & Expansion (Ongoing)
Roll out the solution across your network, ensuring scalability and isolation for multiple concurrent dApps. Monitor performance, fine-tune models, and explore additional AI-native RAN and ISAC use cases to maximize long-term value and evolve with 6G standards.
Ready to Transform Your RAN?
Unlock the full potential of AI-native 6G with GPU-accelerated edge inference.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat