Large Language Models (LLMs) have transformed society, but their computational and energy needs hinder efficient inference. The memory wall, the growing processor-memory speed disparity, remains a critical bottleneck for LLM. While Process-in-Memory (PIM) architectures address this challenge by co-locating computation with memory, achieving 5-20 × higher bandwidth than GPUs, existing scalable PIM solutions face critical trade-offs in flexibility, capacity, and efficiency when handling LLMs' dynamic memory-compute patterns and operator diversity. DRAM-PIM suffers from inter-bank communication overhead despite its vector parallelism. SRAM-PIM offers sub10ns latency for matrix operation but is constrained by limited capacity. This work introduces CompAir, a scalable PIM architecture that integrates DRAM-PIM and SRAM-PIM with hybrid bonding, enabling efficient linear computations while unlocking multi-granularity data pathways. We further develop CompAirNoC, an advanced network-on-chip (NoC) with an embedded arithmetic logic unit that performs non-linear operations during data movement. Such a design offloads the centralized communication bottleneck in the channel level to distributed banks, simultaneously reducing communication overhead and area cost for scalability. Finally, we develop a hierarchical Instruction Set Architecture that ensures both flexibility and programmability of the hybrid PIM. Experiments show CompAir delivers 1.83-7.98× faster prefill and 1.95 - 6.28 × faster decoding versus state-ofthe-art PIM designs, with 3.52 × lower energy than GPU-PIM hybrids. This work presents the first systematic exploration of hybrid DRAM-PIM and SRAM-PIM architectures with innetwork computation, paving the way towards a scalable PIM system for LLM inference.

Training large language models (LLMs) has become increasingly expensive due to the rapid expansion in model size. Pipeline parallelism is a widely used distributed training technique. However, as LLMs with larger context become prevalent and memory optimization techniques advance, traditional PP methods encounter greater communication challenges due to the increased size of activations and gradients of activations. To address this issue, we introduce weight-pipeline parallelism (WeiPipe) that transitions from an activation-passing pipeline to a weight-passing pipeline. WeiPipe reduces communication costs and achieves a more balanced utilization by transmitting only weights and their gradients between workers in a pipeline manner. WeiPipe does not rely on collective communication primitives, thus ensuring scalability. We present four variations of WeiPipe parallelism, including WeiPipe-Interleave, which emphasizes communication efficiency, and WeiPipe-zero-bubble, discussing the potential for minimal bubble ratios. Our implementation of WeiPipe-Interleave, performed on up to 32 GPUs and tested in various model configurations, including large-context LLM training, demonstrates a significant improvement in throughput compared to state-of-the-art pipeline parallelism and fully sharded data parallelism with different underlying infrastructures, including NVLink connections within cluster with Ethernet among cluster, and PCIe within cluster and Ethernet among cluster. Additionally, WeiPipe also shows greater scalability in communication-constrained scenarios compared to state-of-art strategies.

The rapid development of deep learning has propelled many real-world artificial intelligence applications. Many of these applications integrate multiple neural networks (multi-NN) to cater to various functionalities. There are two challenges of multi-NN acceleration: (1) competition for shared resources becomes a bottleneck, and (2) heterogeneous workloads exhibit remarkably different computing-memory characteristics and various synchronization requirements. Therefore, resource isolation and fine-grained resource allocation for each task are two fundamental requirements for multi-NN computing systems. Although a number of multi-NN acceleration technologies have been explored, few can completely fulfill both of these requirements, especially for mobile scenarios. This paper reports a Hierarchical Asynchronous Parallel Model (HASP) to enhance multi-NN performance to meet both requirements. HASP can be implemented on a multicore processor that adopts Multiple Instruction Multiple Data (MIMD) or Single Instruction Multiple Thread (SIMT) architectures, with minor adaptive modification needed. Further, a prototype chip is developed to validate the hardware effectiveness of this design. A corresponding mapping strategy is also developed, allowing the proposed architecture to simultaneously promote resource utilization and throughput. With the same workload, the prototype chip demonstrates 3.62$\boldsymbol{\times}$, and 3.51$\boldsymbol{\times}$higher throughput over Planaria and 8.68$\boldsymbol{\times}$, 2.61$\boldsymbol{\times}$over Jetson AGX Orin for MobileNet-V1 and ResNet50, respectively.

Building hybrid systems that incorporate various processing-in-memory (PIM) devices and processing-near-memory (PNM) technologies can offer complementary advantages in both efficiency and flexibility, while many-core architectures show great potential in deploying data-centric parallel applications with high performance. Compilers for the hybrid PN/IM architecture are critical for enabling such computing systems to be put into practical use. However, most of the existing neural network compilers for PIM or PNM are optimized from the perspective of an operator, and cannot effectively take advantage of a decentralized core-level dataflow with large on-chip memory access bandwidth. Here, we propose a full-stack System-on-graph Compiler (SongC) framework for many-core architecture, which optimizes the efficiency of the PIM devices and leverages the flexibility of the PNM architectures. SongC establishes multi-level graph abstractions to clarify the critical deployment challenges at different levels and generalizes the standard optimizations, decoupling versatile algorithms and diverse types of hardware. To handle the complexity of many-core resource utilization, we also establish a simulation-compilation interaction flow, including a just-in-time evaluator to boost the scheduling search and an extended Roofline model, referred to as the Palace model, to guide the search. Experiments demonstrate the various optimizations and overall performance of SongC and reveal the capability of strategy exploration.