Videos
TOSSI · MEDIA
Videos
Recordings, demos, and walkthroughs from the TOSSI project — end-to-end Open RAN interoperability, testing at scale, and GPU acceleration.
NTN4:26NTN E2E Testbed :Release A - GEO NTN with OCUDU RAN and OAI UE Simulator
NTN Testbed Release A is the initial open-source software implementation of a 5G Non-Terrestrial Network testbed. This demonstration presents a fully functional 5G Standalone network utilizing a simulated 240 millisecond geostationary satellite link, employing OCUDU RAN, an OAI UE Simulator via ZMQ, and a 5G Core all hosted on a single machine without radio frequency hardware. Successful operation was verified through cell search, SIB19 acquisition, random access, RRC connection setup, NAS authentication, PDU session establishment, and end-to-end user plane IP traffic. The current configuration achieves approximately 510 millisecond round-trip latency and validates scheduler improvements for 3GPP Release 17 NTN compatibility; future releases will incorporate low Earth orbit satellite support with variable propagation delays and Doppler shifts.
Integration7:55Jul 28, 2026Closed-loop Cell Congestion Optimization rApp with OCUDU RAN StackThis demonstration presents an end-to-end Cell Congestion Optimization workflow implemented with Open RAN standards and open-source software, including the OCUDU RAN Stack, Aether SD-Core, and ONAP-based SMO. The system monitors O1 performance data to detect congestion, selects a neighboring cell based on configurable policies, and initiates a handover request via the management plane using standardized interfaces like O1, NETCONF/YANG, and RESTCONF. The workflow validates interoperability between independently developed components through a closed-loop control process involving live telemetry collection, handover generation, and execution. This implementation provides a foundation for autonomous network operations and intelligent rApps in evolving Open RAN environments. The complete source code and technical documentation are publicly available.
AI-RAN5:46Jul 26, 2026OCUDU AI-RAN Release 3 : ML-Based CSI Prediction for Intelligent 5G SchedulingOCUDU AI-RAN Release 3 improves downlink link adaptation by predicting channel quality reports instead of relying on periodically received, potentially stale data. The system utilizes two prediction engines, a linear filter and a neural network, to forecast per-user channel trends and inform scheduling decisions. This pipeline compiles a complete stack, collects over-the-air channel data for training, and runs live base stations with the predictive scheduler. Benchmarking demonstrates performance gains compared to conventional methods while maintaining standards compliance through arithmetic operations within the DU without new signaling or external dependencies. The framework offers automatic fallback mechanisms ensuring operational safety.
Digital Twins9:14Jul 21, 2026NVIDIA Sionna RT Meets OCUDU: Building Deterministic Digital Twins for 6G and AI-RANNVIDIA Sionna, an open-source GPU-accelerated platform for AI-native wireless research, has been integrated with the OCUDU RAN ecosystem via its differentiable ray tracer, Sionna RT. This integration enables a live connection between 3D physical geometry and a functional protocol stack through asynchronous channel streaming using ZeroMQ PUB/SUB. The system employs a GPU-based physics engine for radio wave propagation simulation, transferring data to the CPU for deterministic channel filtering with AVX-512 operations, and supports hot-swapping of channel tap sets without connection interruption. This architecture facilitates site-specific physical realism for applications including AI-driven beamforming and localized LLM fine-tuning in telecom environments.
Integration13:20Jul 17, 2026NVIDIA Aerial, xFAPI & OCUDU: Building the Future of Open AI-RANThis demonstration presents xFAPI Release 2.3 facilitating interoperability between NVIDIA Aerial Layer 1 and OCUDU Layer 2 in an end-to-end Open RAN configuration. The setup also includes AERIAL_OAI mode, which integrates NVIDIA Aerial with Duranta (OAI) to demonstrate flexibility across different ecosystems. Key components used in the deployment are NVIDIA Aerial L1, xFAPI Release 2.3, OCUDU L2, LiteON O-RU, Aether SD-Core, a commercial UE, and Netweb Tyrone Spark AI Supercomputer. This integration showcases GPU-accelerated Layer 1 platforms working with open-source Layer 2 ecosystems for practical AI-RAN deployments.
AI-RAN4:32Jul 10, 2026OCUDU AI-RAN: Building an Open Framework for Per-UE IntelligenceAI-RAN Series Release 2 introduces an open framework for machine-learning-based adaptive Buffer Status Report (BSR) configuration in OCUDU RAN. The release lets the RAN learn each individual UE's traffic behaviour, predict its uplink packet arrival pattern, and dynamically adapt the periodicBSR-Timer per UE at runtime. A small learned model (SVR over Random Fourier Features) predicts the next uplink interarrival, snaps it to the nearest legal TS 38.331 timer, and reconfigures the UE over F1AP, trading signalling cost against how quickly the gNB learns of new uplink data. Disabled by default, it falls back cleanly to the static timer.
Integration7:44Jul 7, 2026OCUDU RAN: Standards-Based Inter-gNB Handover with O-RAN O1 ManagementSeamless inter-gNB handover in a real-world Open RAN deployment with OCUDU RAN, SD-Core, LiteON radio units, and commercial 5G UEs. Demonstrates both operator-initiated (gNB-triggered) and UE-initiated (3GPP A3 measurement-based) handovers with uninterrupted mobility across cells, and introduces O-RAN O1 support for handover monitoring and control, standardized mobility KPIs, and a foundation for AI-powered mobility xApps and rApps.
5G Core5:06Jul 4, 2026Introducing Cloud Native Telecom Certification (CNTC): Trusted Certification for 5G UPFsIntroducing Cloud Native Telecom Certification (CNTC), an open, trusted certification process for 5G User Plane Functions. Covers how CNTC validates UPF conformance, performance, and cloud-native readiness so operators can deploy open-source and third-party UPFs with confidence.
AI-RAN3:33Jul 1, 2026OCUDU AI-RAN Release 1: Building an Open Framework for Intelligent Link AdaptationOCUDU AI-RAN Release 1, an open, end-to-end framework for building, deploying, and continuously improving intelligent RAN applications. The first application is ML-based uplink link adaptation, where a gradient-boosted model picks the MCS alongside OLLA, built on 900k+ real over-the-air transmissions.
5G Core4:49Jun 27, 2026From Kernel-Locked to Cloud-Native: Rebuilding Magma’s 5G User Plane with eBPFRebuilding Magma’s 5G user plane from a kernel-locked data path to a cloud-native one with eBPF. Moves packet processing into eBPF/XDP programs for a portable, programmable, high-throughput user plane that runs on commodity Linux without custom kernel modules.
Integration4:21Jun 26, 2026Building an End-to-End Open Network Slicing Ecosystem with OCUDUA complete open-source network slicing ecosystem built around OCUDU. Multi-UE OAI orchestration with independent S-NSSAIs, slice-aware DU scheduling via 3GPP RRM Policy Ratios, Magma Core integration, and O1-based live slice configuration through the SMO, with per-slice performance monitoring for rApp-driven closed-loop automation.
GPU Offload4:32Jun 23, 2026Building the AI-RAN Data Path in OCUDU: Inline GPU Processing for PRACH and SRSInline GPU acceleration in OCUDU Open RAN using NVIDIA GPUDirect RDMA. Fronthaul packets are delivered directly from the NIC into GPU memory where PRACH detection and SRS channel estimation execute inline, with no CPU copy. Shows 3.3× speedup for PRACH (106 µs vs 352 µs) and sub-linear SRS scaling up to 256 UEs.
Integration4:06Jun 21, 2026OAI L1 + xFAPI + OCUDU DU-High | End-to-End Open RAN InteroperabilityEnd-to-end Open RAN interoperability across the FAPI split. OAI L1 (nFAPI PNF) talks to OCUDU DU-High through the xFAPI bridge in OAI_OCUDU mode — nFAPI over P5 SCTP and P7 UDP between hosts, then xSM shared-memory transport into the OCUDU MAC scheduler. Demonstrates a two-host deployment with F1AP up to the CU.
Testing3:07May 25, 2026Scaling RAN Testing: From Single UE to Multi-UE SimulationScaling open RAN testing from a single UE to a multi-UE simulation. Walks through the OAI ZMQ-capable nr-uesoftmodem, a Python IQ-superposition proxy that fans the downlink and sums uplinks for N UEs, per-UE Linux network namespaces, wave admission into the 5G core, and a live dashboard driving the demo.
Integration2:52May 19, 2026xFAPI Demo: Disaggregated FAPI Interface in Open RANA disaggregated FAPI interface in open RAN. Walks through the latest xFAPI release: interoperable deployment of L1 and L2 through a modular FAPI-based architecture and an xSM shared-memory transport.
GPU Offload4:31May 3, 2026GPU Offloading of PRACH Detection Using CUDA GraphsGPU offloading of PRACH detection using CUDA graphs. Walks through NVIDIA GPU acceleration for PRACH preamble detection in the upper-PHY via cuFFTDx and CUDA-graph capture, with A/B performance results measured on live 5G cells against both Split 8 and Split 7.2 radios.
GPU Offload2:49Apr 26, 2026Offloading LDPC in PDSCH and PUSCH to AcceleratorsOffloading LDPC in PDSCH and PUSCH to hardware accelerators. Walks through the integration of accelerator cards (demonstrated on Intel ACC100) for LDPC offload into the upper-PHY via DPDK BBDEV, with A/B performance results on a live 100 MHz TDD cell against a real UE.