Demonstrated on NVIDIA Spark with DeepSig AI-native PHY
Earlier this month at Mobile World Congress, the Linux Foundation launched OCUDU, often described as the “Linux of RAN.” The initiative focuses on rapidly building a carrier-grade, open source Central Unit (CU) and Distributed Unit (DU) software stack that implements the core functions at the heart of modern radio access networks. OCUDU aims to establish a rigorous, vendor-neutral baseline that provides a mature platform for RAN vendors, operators, and researchers to build upon, enabling the ecosystem to more easily customize 5G deployments for a wide range of use cases while accelerating innovation toward AI-RAN and future 6G technologies.
By providing a shared software foundation for key RAN components, OCUDU streamlines integration and interoperability across the ecosystem, allowing organizations to focus on differentiated capabilities and advanced network functionality. In doing so, the project helps strengthen the broader Open RAN landscape while creating a more efficient path for collaboration, innovation, and deployment of next-generation wireless technologies.
Economists have long modeled the game theory of open source ecosystems in terms of where an industry is best off collaborating for a shared-advantage while also competing in features which are differentiating and which drive innovation and value – OCUDU seeks to move this barrier, allowing for greater industry collaboration and both American next-G innovation and global collaboration – by collaborating on larger portions of the stack, enabling innovation and path to market faster from new and innovative solutions which drive value for operators, new private network deployments, use cases like integrated sensing and communication (ISAC), and of course features which department of defense and federal deployments need to maximize their combined communications, sensing, and operations.
OUSD R&E’s Future G office was instrumental in bootstrapping the OCUDU ecosystem, engaging Software Radio Systems (SRS) and DeepSig to leverage a proven foundation for 5G software, hardware acceleration, and AI-RAN – and working with Linux Foundation to set up a strong team of founding members including numerous industry veterans, major operators, and research institutes alike – including AT&T, AMD, DeepSig, Ericsson, Nokia, NVIDIA, Softbank, SRS, Verizon, and many more – aligned to establish a powerful foundation which will accelerate 5G and 6G innovation and unlock new performance optimizations and RAN application value – transitioning to a self-sustaining ecosystem fund over the next three years and establishing a strong vendor neutral governance, seeking to speed network technology adoption and offerings to operators.
Linux Foundation has helped set up the proven ecosystem and project governance which helps to form a technocracy, and move beyond political and corporate differences to advance the project at a rapid pace – while also embracing state of the art software security validation, and certificate of origin management to ensure a secure and trusted software supply chain and best practice open source software review, verification and security.
OCUDU is about moving fast and will be driven by technical contributors
While OCUDU is only at its formation, it is built on a mature SRS 5G stack, which is already used in numerous stable and long-running commercial network deployments, has been validated with dozens of radio-unit solutions, has been heavily optimized and tested on both Intel and ARM platforms, and has been rigorously tested in numerous labs including with Deutsche Telekom in the i14y lab, where it was certified for a number of key network functions. DeepSig has, over recent months, accelerated the project by expanding the ability of OCUDU to take advantage of NVIDIA GPU compute platforms such as NVIDIA DGX Spark, NVIDIA Jetson Thor, and NVIDIA GH200 Grace Hopper Superchip, leveraging both OCUDU and NVIDIA AI Aerial in an attempt to bring open source accelerated compute into projects across a broad range of data-center class and edge-deployment class platforms. While OCUDU seeks to support multiple vendors and compute platforms, NVIDIA GPUs have been central to the recent AI ecosystem explosion, at the heart of countless innovations, and a critical platform for combining both efficient RAN compute and AI-RAN applications, within one highly optimized shared memory framework – a powerful platform on which OCUDU can be developed, demonstrated, and deployed.
Demonstrating AI-RAN on Accelerated OCUDU at NVIDIA GTC
DeepSig has adopted OCUDU as the central reference project for AI-RAN, bringing core work in OmniPHY for AI-for-RAN air-interface optimization, and OmniSIG spectrum awareness work onto the platform, unleashing efficiencies in 5G as well as 6G prototypes and sensing applications beyond the core communications mission onto one platform, which is needed in both commercial and defense applications. At GTC, we showcase OCUDU running on a highly optimized DGX Spark platform, running both RAN and AI-RAN functions efficiently on an embedded shared memory ARM and GPU platform, which provides a rapid path to adoption of a wide range of key next-G use cases. Specifically, we showcase a full 5G stack with both transparent AI-RAN in the form of neural receivers, improving coverage, capacity, and throughput for currency day standards, alongside two-sided AI models which realize the vision for 6G AI-Native air interfaces, seamlessly operating on the same sector without hardware modification, as well as AI-Native spectrum sensing and awareness from the RAN, critical to societal and national security as a key RAN function.
This highlights key work from DeepSig’s AI-RAN Alliance work item #1, the effort on a learned AI-native air interface, as well as other critical AI-RAN capabilities running alongside.
Building the Foundations for OCUDU and 6G RAN
OCUDU provides rigorously tested, performance-oriented open-source implementations of critical CU and DU components of a 5G base station on multi-vendor hardware platforms. Expanding this to provide broad NVIDIA GPU acceleration support and AI-RAN reference designs helps to drastically accelerate research, prototyping, and a streamlined path to efficient deployment by a broad range of researchers and new and established vendors across a unified ecosystem.
In the image above, at MWC, we showcased an NVIDIA DGX Spark-based gNB that is fully GPU-accelerated in L1 processing, supports over 3 sectors of 100 MHz 4- or 8-TR connectivity, and offers multiple AI-RAN capabilities, all accelerated fluidly within the RAN. These include leveraging a Neural Receiver, which transparently improves uplink performance in 5G compliant receivers, as well as a prototype AI-Native Air Interface, which demonstrates novel 6G prototype physical layer capabilities in line with 3GPP release 20 study items, alongside AI-Native spectrum sensing to understand interference, jamming, and threats in the same band.
While OCUDU has been an enormous enabler of rapid integration, verification, and deployment of combined AI-RAN capabilities, this is only the beginning. DeepSig’s commitment to the OCUDU ecosystem will see the open sourcing of GPU-accelerated core L1 processing functions within the coming weeks, as well as the publication of reference designs, benchmarks, and exemplars, including AI-RAN reference designs, within the next few months to turbocharge RAN innovation and performance.
These capabilities will help to form the shared and interoperable foundation of the next generation of intelligent networks – driving the most critical 5G and 6G RAN performance and the critical AI-RAN performance integrally linked with them on the same platform – while providing multi-vendor reference platforms which will help drive energy and cost efficiency across a broad range of operators, deployments, and use cases.
Attending NVIDIA GTC? Visit the AI-RAN Alliance booth to see a compact DGX Spark-powered AI-RAN base station running 5G and a prototype “6G” MRSS (transparent joint operation on a single cell) live over the air, highlighting this work on top of the OCUDU platform.
