Energy efficiency and real-time performance are among the main challenges in the 5G/6G era, driven by the increasing complexity of computational algorithms in stream processing and AI-based applications. To meet these demands, computing is shifting toward being distributed and executed on-the-fly as data flows through network elements, a paradigm referred to as In-Network Computing (INC). While initial demonstrations have shown that small kernels can be offloaded to programmable network devices, deploying multi-kernel applications remains difficult due to constrained memory and compute resources, limited floating-point support, and the complexity of handling heterogeneous operations. Addressing these challenges requires new methods that jointly consider application characteristics, approximation opportunities, and hardware/software limitations across the edge-to-cloud continuum.

The X-DNet project tackles these challenges through a cross-layer methodology and architecture that integrates various approximation and optimization techniques along with adaptive and reconfigurable accelerator design. First, we reduce application complexity by exploiting approximation opportunities in data acquisition and computation phases. Second, we design and configure accelerators for heterogeneous network elements, simplifying complex arithmetic operations to match the capabilities of programmable switches. Third, we conduct application-level sensitivity analyses to explore the trade-offs between performance and quality-of-results (QoR), using a greedy heuristic to efficiently generate Pareto/near-Pareto mixed-precision configurations. Together, these contributions enable dynamic deployment of multi-kernel applications in INC environments, delivering improved performance and energy efficiency while maintaining user-defined QoR requirements.

1. Prof. Dr. Akash Kumar (head of supervision)
2. Dr-Ing. Zahra Ebrahimi (project manager and research associate)
3. M.Sc. Maryam Eslami (research associate)

1. Zahra Ebrahimi, Maryam Eslami, Xun Xiao, and Akash Kumar. "X-DINC: Toward Cross-Layer Approximation for the Distributed and In-Network Acceleration of Multi-Kernel Applications" Elsevier Future Generation Computer Systems, Volume 172, 2025.

Project website

Energy efficiency and real-time performance are among the main challenges in the 5G/6G era, driven by the increasing complexity of computational algorithms in stream processing and AI-based applications. To meet these demands, computing is shifting toward being distributed and executed on-the-fly as data flows through network elements, a paradigm referred to as In-Network Computing (INC). While initial demonstrations have shown that small kernels can be offloaded to programmable network devices, deploying multi-kernel applications remains difficult due to constrained memory and compute resources, limited floating-point support, and the complexity of handling heterogeneous operations. Addressing these challenges requires new methods that jointly consider application characteristics, approximation opportunities, and hardware/software limitations across the edge-to-cloud continuum.

The X-DNet project tackles these challenges through a cross-layer methodology and architecture that integrates various approximation and optimization techniques along with adaptive and reconfigurable accelerator design. First, we reduce application complexity by exploiting approximation opportunities in data acquisition and computation phases. Second, we design and configure accelerators for heterogeneous network elements, simplifying complex arithmetic operations to match the capabilities of programmable switches. Third, we conduct application-level sensitivity analyses to explore the trade-offs between performance and quality-of-results (QoR), using a greedy heuristic to efficiently generate Pareto/near-Pareto mixed-precision configurations. Together, these contributions enable dynamic deployment of multi-kernel applications in INC environments, delivering improved performance and energy efficiency while maintaining user-defined QoR requirements.

1. Prof. Dr. Akash Kumar (head of supervision)
2. Dr-Ing. Zahra Ebrahimi (project manager and research associate)
3. M.Sc. Maryam Eslami (research associate)

1. Zahra Ebrahimi, Maryam Eslami, Xun Xiao, and Akash Kumar. "X-DINC: Toward Cross-Layer Approximation for the Distributed and In-Network Acceleration of Multi-Kernel Applications" Elsevier Future Generation Computer Systems, Volume 172, 2025.

Project website