Network acceleration for generative AI clusters involves optimizing the communication and data transfer between multiple AI models, often distributed across different compute nodes. In this context, a large number of research challenges need to be tackled. Reducing latency is crucial to support model parallelism and real-time interactions with end-users. Controlling how bandwidth is utilized is key to maximize throughput (large amount of data can be quickly transferred) and minimize job completion times. As clusters grow in size, scalability is also a challenge as the complexity of managing communication increases. Ensuring that the network can handle node failures gracefully is essential for reliability and fault tolerance. Optimized software frameworks using AI frameworks that are optimized for specific hardware accelerators, lead to better performance. Large Language Models (LLM) are so large that a dedicated computing cluster with thousands of GPUs is now needed both at training and inference time. E.g., Meta is designing dedicated clusters with the goal of training LLM in 1 day instead of 1 month. To accommodate ever-larger models, data parallelism is no longer enough and there has been a recent shift towards more sophisticated model parallelism techniques. Large models are divided into smaller, manageable pieces that can be distributed across multiple GPUs or TPUs. As a large number of messages, of small size, together with massive sets of elephants flows are exchanged, efficient communication between devices is essential. Latency builds up very quickly in large clusters, along with the number of hops, and potential congestion. The network fabric is becoming a significant bottleneck: network acceleration techniques will be developed from in Net4AI to remove this limit and deeply evaluated according to acceleration and power efficiency.

The LeadingEdge project will deliver a novel and holistic framework to efficiently cope with unresolved challenges in edge computing ecosystems, regarding dynamic resource provisioning to multiple coexisting services amidst unknown service- and system-level dynamics. The project approach is three-faceted; it will optimize intra-service resource provisioning, inter-service resource coordination, and user perceived quality of experience (QoE). First, at service level, we will develop a framework, grounded on first principles, for opportunistic use of edge and cloud computation, bandwidth and cache resources according to instantaneous resource availability, mobility, connectivity, service resource requirements and service demand. Our approach will rely on solid online-learning theories such as online convex optimization (OCO), and transfer learning and it will eliminate our inherent inability to predict demand, mobility, and other dynamic processes that affect resource allocation. It will also use extreme-value theory and stochastic optimization towards a full-fledged study of the latency-reliability trade-off that is fundamental for mission-critical services. Proof-of-concept (PoC) validation will be provided through, (i) a real-time image recognition tool as part of a video analytics procedure, (ii) two alternative video quality assessment solutions with different degree of complexity and different configurations of edge/client or cloud resources. After service-level optimization, at a second level, we will develop a system-level AI-empowered service orchestrator based on reinforcement learning and context awareness for service orchestration in terms of network slicing and service chain placement, such that instantaneous service-level requirements are fulfilled. The OpenAirInterface.org (OAI) and Mosaic-5g.io software platforms will be used as real-time experimentation environments with full 4G/5G functionalities for service orchestration to place services, direct traffic from users to servers, and measure latency and other QoE metrics. Finally, at user level, we will leverage the community-network infrastructure of guifi.net as an edge network to deploy services at scale in a controlled manner and to directly measure their impact on user QoE. The outcome of these latter user-level studies will be continually fed back to and guide the service- and the system-level optimization. The project results are envisioned to be transformational for edge computing and to create durable impact through enabling game-changing services. This ambitious objective will be pursued with a balanced consortium of complementary expertise, consisting of 3 universities, a research centre, a SME, and a large industry, overall spanning 4 countries.

DYNAMOS develops fast (1 ns) and widely tunable (>110 nm) lasers, energy-efficient (~ fJ/bit), broadband (100 GHz) electro-optic modulators, and high-speed (1 ns) broadcast-and-select packet switches as photonic integrated circuits (PICs). DYNAMOS meets the expected outcome objectives and call scope by proposing the development of low energy (few pJ/bit) PICs, which are integrated into modular and scalable subsystems, and subsequently utilized to demonstrate novel data centre networks with highly deterministic sub-microsecond latency to enable maximum congestion reduction, full bisection bandwidth (lower congestion) and guaranteed quality of service while reducing cost per Gbps. The proposed network offers optical circuit switched reconfiguration and guaranteed (contention-less) full-bisection bandwidth, allowing any computational node to communicate to any other node at full-capacity. DYNAMOS builds on recent developments in III-V optoelectronics, thick silicon-on-insulator waveguide technology, and silicon organic hybrid (SOH) modulators. It co-develops the entire ecosystem of transceivers, switches and networks to boost overall performance and to reducing the total cost of data exchange, instead of focusing on the improvement of individual optical links or interfaces. The objectives of DYNAMOS perfectly match the major photonics research & innovations challenges defined in the Photonics21 Multiannual Strategic Roadmap 2021-2027.

The burden of cardiovascular disease (CVD) on society is huge with >85 million people affected in Europe. The overall prevalence continues to grow due to unhealthy lifestyles and population aging. Heart failure (HF) is the final common pathway of all CVD and has a 5 year mortality rate of 20-50% despite significant advances in therapy. iCARE4CVD aims to address this burden by contributing to three essential steps to improve the current care pathways, covering all stages from early risk to established HF: 1) early diagnosis to identify patients at risk of CVD and divide them into clinically meaningful subgroups; 2) risk stratification for these subgroups to define the urgency for intervention; and 3) prediction of treatment response for each subgroup. This will be achieved by the following steps: clinical partners will provide a large set of cohorts including >1,000,000 patients with a wide range of biomarkers (e.g. digital, blood, imaging). Anonymous access to data will be enabled by using a blockchain-supported federated database. Artificial intelligence-based modeling also considering patient relevant factors will assess changes in risk and stratify patients according to their individual responses to therapy. Results will then be prospectively validated in new and ongoing large cohorts and a pilot trial to test the prediction of treatment response by using multiple biomarkers going beyond current risk prediction (such as SCORE) towards individualized therapy. Results will be used to provide novel decision tools for each step targeting newly identified subgroups and as a blueprint for innovative future trials to individualise prevention and therapy. Patient involvement is key in every part of iCARE4CVD (e.g. patient advisory board, country-specific Patient Panels) to build a motivational framework for self-care by patients. The project brings together an EU-wide consortium with the needed resources and expertise from the public and private side to bring iCARE4CVD to success.