Technology / AI-RAN
Accelerated RAN on Intel and NVIDIA - radio and AI on one platform.
AiraGrid builds the radio access network on current, accelerated silicon: Intel Xeon 6 with integrated vRAN Boost for single-server cell sites, and NVIDIA Aerial CUDA-accelerated O-RAN for the AI-RAN path.
Intel accelerated RAN
One rugged box at the cell site, on Xeon 6.
The AiraEdge baseband runs on the Intel Xeon 6 SoC (Granite Rapids-D). Integrated 100/200G Ethernet and VRB2 vRAN Boost fold the layer-1 accelerator, the NICs and the host CPU into a single die - replacing a rack of appliances with one fanless, far-edge server.
Xeon 6 SoC (Granite Rapids-D)
A far-edge vRAN SoC with integrated 100/200G Ethernet and on-die acceleration - collapsing a multi-server cell site into a single rugged box.
Integrated vRAN Boost (VRB2)
FEC and FFT acceleration built into the Xeon die, removing the need for a separate ACC100 card while keeping the layer-1 pipeline on x86.
FlexRAN reference PHY
Intel's FlexRAN PHY - the validated x86 software stack proven across the private-wireless industry - powers the AiraEdge baseband.
On-die AI with AMX
Advanced Matrix Extensions bring inference acceleration onto the same CPU that runs the RAN, with no discrete accelerator at the site.
Tiber Open Edge orchestration
Intel's Open Edge Platform brings zero-touch provisioning and fleet orchestration to distributed far-edge nodes - so a network of single-server sites is managed as one fleet, not box by box.
- Compute
- Intel Xeon 6 SoC (Granite Rapids-D), up to 72 cores
- L1 acceleration
- Integrated vRAN Boost (VRB2): FEC + FFT on-die, no ACC100 card
- Networking
- Integrated 100/200G Ethernet on-package
- Software PHY
- Intel FlexRAN reference PHY (x86)
- On-die AI
- Advanced Matrix Extensions (AMX)
- Form factor
- Fanless / rugged, single-server far-edge cell site
NVIDIA AI-RAN
When the site needs AI, the RAN already has the accelerator.
For sites that pair connectivity with on-site intelligence, the AiraEdge AI-RAN SKU runs NVIDIA Aerial - an O-RAN 7.2x O-DU on GPU - alongside AI workloads on the same accelerated platform. ARC-Compact (Grace C1 + L4) provides the reference node; ARC-1 (GB200 NVL2) scales it up for aggregation and centralized RAN.
Aerial CUDA-accelerated RAN
A software-defined, O-RAN 7.2x-compliant O-DU running on GPU - built so RAN and AI workloads share one accelerated platform.
ARC-Compact (Grace + L4)
The Aerial RAN Computer reference hardware pairs an Arm-efficient Grace C1 host with an L4 GPU in a compact 2RU node for cell-site deployment.
Scales to aggregation
ARC-Compact (Grace C1 + L4) is the reference cell-site node; ARC-1 (GB200 NVL2) scales the same Aerial platform up for aggregation and centralized RAN.
AI-RAN on one accelerator
When radio load is light, the same GPU runs on-prem inference instead of sitting idle. One box at the site does both, collapsing the cost, latency and footprint that two separate stacks would carry.
- O-DU
- NVIDIA Aerial CUDA-accelerated RAN, O-RAN 7.2x split, software-defined
- Reference node (D-RAN)
- ARC-Compact: Grace C1 + L4 GPU, 2RU, -5 C to +55 C
- Aggregation (C-RAN)
- ARC-1 (GB200 NVL2)
- Host CPU
- NVIDIA Grace C1 (Arm-efficient, 140-250 W)
- Shared infrastructure
- RAN + AI on one accelerated platform (AI-RAN)
The same accelerated thinking carries into the core: AiraOne runs cloud-native CNFs on bare-metal Xeon 6 via Intel's BMRA, so RAN and core share one silicon philosophy end to end.
Carrier-grade by design
Accelerated silicon is the engine. These make it a network.
Deterministic and URLLC
Tuned for ultra-reliable low-latency control, the deterministic fabric that autonomous fleets, crane control and remote operation depend on.
Network slicing
Per-application slices isolate traffic and hold an SLA per service on the same private network, so safety control and bulk video never contend.
Timing and sync
PTP and GNSS timing with holdover keep the radio network precisely synchronized for TDD and coordinated, time-sensitive control.
RIC-driven optimization
A RAN Intelligent Controller - near-real-time and non-real-time RIC running xApps and rApps - steers traffic, saves energy and flags anomalies: the AI in AI-RAN working on the network itself, not only the applications.
The principle
Build forward, not on hardware two generations old.
AiraGrid standardises on silicon the vendors are actively roadmapping - Xeon 6 / Granite Rapids-D, Aerial, Grace and Orin - so the network has headroom for AI from day one and a clear path to forward generations. The AiraEdge AI-RAN SKU is built on NVIDIA's MGX modular server reference architecture, so accelerated-RAN nodes are productized fast on validated, supportable hardware rather than bespoke one-offs.
Today's builds ship on Orin, Grace and GH200; the roadmap runs to GB200 / Vera and IGX Thor as they reach volume, stated as roadmap, never sold as shipping.
Bring a private network to your toughest site.
Talk to our team about a deployment built for your terrain, your spectrum and your operations.