MEDEM: Multi-Engine DL Accelerator Design Methodology

MEDEM (Multi-Engine DL Accelerator Design Methodology), published on arXiv:2609.37399v1 on September 29, 2026, is a systematic two-stage methodology for designing flexible multi-engine deep learning accelerators. It decouples engine co-design from combination selection to efficiently navigate exponentially large design spaces and minimize aggregate execution costs.

The methodology employs two core strategies: the Hierarchical Co-designer (HCo), which prioritizes high-impact hardware-mapping axes for exhaustive exploration, and Balanced Average and Specialization Search (BASS), which selects engine multisets that balance common-case replication with specialized heterogeneity. Comprehensive evaluations using 51 DL workloads demonstrate that MEDEM identifies accelerators outperforming state-of-the-art designs, achieving geometric-mean improvements of up to 4.84× in energy-delay product (EDP) and 1.59× in throughput.

MEDEM: Multi-Engine DL Accelerator Design Methodology addresses the architectural challenge of designing deep learning accelerators that integrate multiple specialized compute engines rather than relying on a single general-purpose datapath. As modern DL workloads become increasingly heterogeneous—spanning convolutional, attention-based, memory-bound, and irregular kernels—no single engine type can simultaneously optimize latency, throughput, energy efficiency, and utilization. The paper frames this as a design-space problem: an accelerator must select and combine engines with complementary computational characteristics so that different workload components can be matched to the engines best suited to their arithmetic intensity, data-movement patterns, and precision requirements.

Methodologically, MEDEM contributes a structured approach for composing such multi-engine architectures, moving accelerator design away from ad hoc engine selection toward a more systematic evaluation of workload demands, engine capabilities, and system-level integration. The central insight is that efficiency comes not only from building a fast core, but from aligning a heterogeneous set of engines with the distinct phases and operators of DL workloads, while accounting for the trade-offs introduced by multi-engine coordination. This makes the methodology useful for architects who need to balance performance, energy efficiency, flexibility, and implementation complexity in next-generation AI hardware.

The work matters because multi-engine DL accelerators are becoming a practical path to serving diverse workloads on shared silicon, especially in edge, datacenter, and heterogeneous SoC environments where a single fixed engine would often be underutilized or inefficient. By providing a design methodology, MEDEM helps reduce the search space and clarifies the trade-offs involved in choosing engine mixes, potentially leading to more predictable performance, better cost/energy outcomes, and more adaptable accelerator designs for evolving DL workloads.