Domain-Specific Computer Architectures for Emerging Applications II

Domain-Specific Computer Architectures for Emerging Applications II: From Deep Learning to Large Language Models provides a systematic account of the architectural shift from task-specific deep learning accelerators to the computing platforms required by transformer-based large language models, Vision Transformers, and mixture-of-experts networks.

As AI models scale, performance is determined not only by arithmetic throughput but also by memory bandwidth, data movement, communication efficiency, compiler support, and full-stack hardware-software co-design. This book explains why acceleration strategies developed for convolutional neural networks are no longer sufficient for many contemporary workloads, and presents the architectural principles required for self-attention, sparse execution, KV-cache management, heterogeneous acceleration, and cluster-scale inference. By connecting algorithmic structure with accelerator design, compiler automation, and distributed systems, it offers a unified technical framework for modern AI computing. Topics covered include GPUs, FPGAs, ASICs, spatial accelerators, sparse tensor compilation, auto-tuning, high-level synthesis, neural architecture search, and scalable large language model serving.

Combining conceptual foundations with concrete system methodologies, the book is intended for graduate students, researchers, and practitioners in computer architecture, AI systems, and hardware-software co-design seeking to understand how specialized computing platforms are evolving for the foundation-model era.