Transformer-based models are the foundation of modern machine learning, but their execution, particularly during autoregressive decoding in large language models (LLMs), places significant pressure on memory systems due to frequent memory accesses and growing key-value (KV) caches. This creates a bottleneck in memory bandwidth, especially as context lengths increase. Processing-in-memory (PIM) architectures are a promising solution, offering high internal bandwidth and compute parallelism near memory. However, current PIM designs are primarily optimized for dense attention and struggle with the dynamic, irregular access patterns introduced by modern KV cache sparsity techniques. Consequently, they suffer from workload imbalance, reducing throughput and resource utilization. In this work, we propose STARC, a novel sparsity-optimized data mapping scheme tailored specifically for efficient LLM decoding on PIM architectures. STARC clusters KV pairs by semantic similarity and maps them to contiguous memory regions aligned with PIM bank structures. During decoding, queries retrieve relevant tokens at cluster granularity by matching against precomputed centroids, enabling selective attention and parallel processing without frequent reclustering or data movement overhead. Experiments on the HBM-PIM system show that, compared to common token-wise sparsity methods, STARC reduces attention-layer latency by 19%--31% and energy consumption by 19%--27%. Under a KV cache budget of 1024, it achieves up to 54%--74% latency reduction and 45%--67% energy reduction compared to full KV cache retrieval. Meanwhile, STARC maintains model accuracy comparable to state-of-the-art sparse attention methods, demonstrating its effectiveness in enabling efficient and hardware-friendly long-context LLM inference on PIM architectures.

Recent advances in machine learning (ML) have spotlighted the pressing need for computing architectures that bridge the gap between memory bandwidth and processing power. The advent of deep neural networks has pushed traditional Von Neumann architectures to their limits due to the high latency and energy consumption costs associated with data movement between the processor and memory for these workloads. One of the solutions to overcome this bottleneck is to perform computation within the main memory through processing-in-memory (PIM), thereby limiting data movement and the costs associated with it. However, DRAM-based PIM struggles to achieve high throughput and energy efficiency due to internal data movement bottlenecks and the need for frequent refresh operations. In this work, we introduce OPIMA, a PIM-based ML accelerator, architected within an optical main memory. OPIMA has been designed to leverage the inherent massive parallelism within main memory while performing high-speed, low-energy optical computation to accelerate ML models based on convolutional neural networks. We present a comprehensive analysis of OPIMA to guide design choices and operational mechanisms. Additionally, we evaluate the performance and energy consumption of OPIMA, comparing it with conventional electronic computing systems and emerging photonic PIM architectures. The experimental results show that OPIMA can achieve 2.98x higher throughput and 137x better energy efficiency than the best-known prior work.

Machine Learning (ML) training on large-scale datasets is a very expensive and time-consuming workload. Processor-centric architectures (e.g., CPU, GPU) commonly used for modern ML training workloads are limited by the data movement bottleneck, i.e., due to repeatedly accessing the training dataset. As a result, processor-centric systems suffer from performance degradation and high energy consumption. Processing-In-Memory (PIM) is a promising solution to alleviate the data movement bottleneck by placing the computation mechanisms inside or near memory. Our goal is to understand the capabilities and characteristics of popular distributed optimization algorithms on real-world PIM architectures to accelerate data-intensive ML training workloads. To this end, we 1) implement several representative centralized distributed optimization algorithms on UPMEM's real-world general-purpose PIM system, 2) rigorously evaluate these algorithms for ML training on large-scale datasets in terms of performance, accuracy, and scalability, 3) compare to conventional CPU and GPU baselines, and 4) discuss implications for future PIM hardware and the need to shift to an algorithm-hardware codesign perspective to accommodate decentralized distributed optimization algorithms. Our results demonstrate three major findings: 1) Modern general-purpose PIM architectures can be a viable alternative to state-of-the-art CPUs and GPUs for many memory-bound ML training workloads, when operations and datatypes are natively supported by PIM hardware, 2) the importance of carefully choosing the optimization algorithm that best fit PIM, and 3) contrary to popular belief, contemporary PIM architectures do not scale approximately linearly with the number of nodes for many data-intensive ML training workloads. To facilitate future research, we aim to open-source our complete codebase.

Neural networks (NNs) are growing in importance and complexity. A neural network's performance (and energy efficiency) can be bound either by computation or memory resources. The processing-in-memory (PIM) paradigm, where computation is placed near or within memory arrays, is a viable solution to accelerate memory-bound NNs. However, PIM architectures vary in form, where different PIM approaches lead to different trade-offs. Our goal is to analyze, discuss, and contrast DRAM-based PIM architectures for NN performance and energy efficiency. To do so, we analyze three state-of-the-art PIM architectures: (1) UPMEM, which integrates processors and DRAM arrays into a single 2D chip; (2) Mensa, a 3D-stack-based PIM architecture tailored for edge devices; and (3) SIMDRAM, which uses the analog principles of DRAM to execute bit-serial operations. Our analysis reveals that PIM greatly benefits memory-bound NNs: (1) UPMEM provides 23x the performance of a high-end GPU when the GPU requires memory oversubscription for a general matrix-vector multiplication kernel; (2) Mensa improves energy efficiency and throughput by 3.0x and 3.1x over the Google Edge TPU for 24 Google edge NN models; and (3) SIMDRAM outperforms a CPU/GPU by 16.7x/1.4x for three binary NNs. We conclude that the ideal PIM architecture for NN models depends on a model's distinct attributes, due to the inherent architectural design choices.

Training machine learning (ML) algorithms is a computationally intensive process, which is frequently memory-bound due to repeatedly accessing large training datasets. As a result, processor-centric systems (e.g., CPU, GPU) suffer from costly data movement between memory units and processing units, which consumes large amounts of energy and execution cycles. Memory-centric computing systems, i.e., with processing-in-memory (PIM) capabilities, can alleviate this data movement bottleneck. Our goal is to understand the potential of modern general-purpose PIM architectures to accelerate ML training. To do so, we (1) implement several representative classic ML algorithms (namely, linear regression, logistic regression, decision tree, K-Means clustering) on a real-world general-purpose PIM architecture, (2) rigorously evaluate and characterize them in terms of accuracy, performance and scaling, and (3) compare to their counterpart implementations on CPU and GPU.