Deep learning models have become foundational across a wide range of applications, demonstrating remarkable performance in diverse tasks. Early deep neural networks (DNNs) primarily comprised dense layers such as fully connected and convolutional laye...
Deep learning models have become foundational across a wide range of applications, demonstrating remarkable performance in diverse tasks. Early deep neural networks (DNNs) primarily comprised dense layers such as fully connected and convolutional layers, which were compute-intensive and well supported by traditional hardware accelerators (e.g., GPUs, TPUs). However, the landscape of deep learning has rapidly evolved with the emergence of large-scale models such as Large Language Models (LLMs), recommender systems (RecSys), and graph neural networks (GNNs). These models introduce new challenges as they demand significantly higher memory bandwidth, larger memory capacity, and improved energy efficiency due to their large parameter sizes and irregular memory access patterns. This dissertation addresses the critical challenge of accelerating memory-intensive operations in deep learning applications by designing hardware architectures tailored to the specific characteristics of each workload.
First, we propose Duplex, a computing architecture that targets the inference of recent LLMs. Duplex integrates two types of processing units within a single device: xPU, which is optimized for high arithmetic intensity (Op/B) operations, and Logic-PIM, a novel logic-die-based processing-in-memory (PIM) architecture designed to handle low-Op/B operations efficiently. Based on workload characteristics, Duplex dynamically selects the appropriate compute unit for each layer. Logic-PIM takes advantage of recent trends in high-bandwidth memory (HBM) technology by placing powerful computing units on the logic die and increasing the number of through-silicon vias (TSVs) between the DRAM dies and the logic die, enabling high-bandwidth between the DRAM dies and computing units. To further improve utilization, we introduce expert and attention co-processing mechanisms, allowing fine-grained parallel execution of MoE and attention layers across both compute units. Evaluation results show that Duplex achieves up to 2.67X higher throughput and 42.0% lower energy consumption than baseline GPU-only systems, with average improvements of 2.07X in throughput and 28.2% in energy efficiency for real-world LLM inference workloads.
Second, we present CLAY, a Compute Express Link (CXL)-based scalable near-data processing (NDP) architecture that accelerates the execution of embedding layers, which are a dominant component in recommendation models and GNNs. Unlike conventional DIMM-based NDP architectures that suffer from memory channel limitations and performance bottlenecks caused by multi-drop bus structures, CLAY introduces a new architecture composed of DRAM clusters with on-board interconnection and dedicated compute units. CLAY significantly reduces data transfer time by allowing direct communication between memory modules and reducing instruction bandwidth by employing packet duplication. Furthermore, we design a fine-grained memory address mapping strategy to minimize load imbalance during sparse memory access of the embedding layers. We extend CLAY to multi-device configurations via a CXL switch, enabling efficient processing of extremely large embedding tables. Experimental results show that CLAY achieves up to 2.77X end-to-end speedup compared to state-of-the-art NDP architectures such as FeaNMP and G-NMP, with substantial gains across both RecSys and GNN.
Together, these contributions highlight the importance of workload-specific architectural specialization and memory-centric design in overcoming performance bottlenecks in modern memory-intensive deep learning applications.