Deploying AI-RAN at the Edge: 6WIND’s Distributed UPF on NVIDIA Grace Hopper
As the telecom industry transitions into the AI era, deploying AI-native workloads at the edge is no longer a futuristic concept. At the heart of this evolution is the combination of 6WIND’s high-performance, software-based User Plane Function (UPF) and NVIDIA’s Grace Hopper platform—a collaboration that takes AI-RAN a significant step closer to real-world deployment, bridging the gap between concept and operational feasibility.
Why AI-RAN Needs a Distributed UPF
AI-RAN architectures demand real-time responsiveness and localized compute. That means user data traffic must be processed close to where it’s generated—at the network edge. Traditional centralized UPFs can’t meet this requirement due to inherent latency and scalability constraints.
This is where 6WIND’s Distributed UPF (dUPF) makes the difference:
- It pushes packet processing to the edge
- It enables low-latency routing for AI applications
- And it reduces reliance on centralized infrastructure, cutting both OPEX and backhaul congestion
Why NVIDIA Grace Hopper Is the Right Fit for AI-RAN Evolution
The NVIDIA Grace Hopper Superchip combines CPU and GPU processing on a single module, purpose-built for high-performance AI and telecom workloads. Its key strengths:
- High memory bandwidth
- Scalable compute performance
- Efficient workload acceleration and power usage
This makes it an ideal host platform for edge workloads like dUPF and AI inferencing, all in one compact, energy-efficient node.
Putting It All Together: AI at the Edge in Action
In a joint performance demo, 6WIND deployed its dUPF on NVIDIA’s Grace Hopper platform to emulate a full AI-RAN use case:
- 30,000 simulated user devices, across 20 gNBs
- 100 Gbps traffic load across two PDU sessions per device
- Local breakout for AI data processing at the edge
All of this was achieved using:
- Only 4 data plane CPU cores
- Less than 80W of power
- With 25µs latency and zero packet loss
This setup demonstrates real-world feasibility: a distributed, AI-ready network that is scalable, deterministic, and resource-efficient.
The Value for Operators
For operators looking to support agentic AI, autonomous systems, and ultra-low latency applications, this architecture :
- ensures edge inferencing without centralized delays
- minimizes jitter (as low as 8µs average)
- reduces OPEX and TCO through efficient hardware usage
- And opens new revenue models for AI-based services
Conclusion
The integration of 6WIND’s distributed UPF with NVIDIA’s Grace Hopper platform enables a practical path to AI-RAN deployment. It delivers not only the performance needed for real-time AI workloads, but also the efficiency and flexibility required for scalable, distributed network infrastructure.
As networks evolve toward 6G and edge-native applications, this solution sets a blueprint for how service providers can meet the demands of AI at scale.
Get the Full Story
Want the technical details, architecture diagrams, and benchmark results?
Download the full white paper here and see how 6WIND and NVIDIA are shaping the future of AI-powered telecom networks.