Large language models (LLMs) have emerged as powerful tools for natural language processing tasks, revolutionizing the field with their ability to understand and generate human-like text. As the demand for more sophisticated LLMs continues to grow, there is a pressing need to address the computational challenges associated with their scale and complexity. This paper presents a comprehensive survey of hardware accelerators designed to enhance the performance and energy efficiency of large language models. By examining a diverse range of accelerators, including GPUs, FPGAs, and custom-designed architectures, we explore the landscape of hardware solutions tailored to meet the unique computational demands of LLMs. The survey encompasses an in-depth analysis of architecture, performance metrics, and energy efficiency considerations, providing valuable insights for researchers, engineers, and decision-makers aiming to optimize the deployment of LLMs in real-world applications.

Optics InfoBase is the Optical Society's online library for flagship journals, partnered and copublished journals, and recent proceedings from OSA conferences.

Emerging web applications like cloud computing, Big Data and social networks have created the need for powerful centres hosting hundreds of thousands of servers. Currently, the data centres are based on general purpose processors that provide high flexibility buts lack the energy efficiency of customized accelerators. VINEYARD aims to develop an integrated platform for energy-efficient data centres based on new servers with novel, coarse-grain and fine-grain, programmable hardware accelerators. It will, also, build a high-level programming framework for allowing end-users to seamlessly utilize these accelerators in heterogeneous computing systems by employing typical data-centre programming frameworks (e.g. MapReduce, Storm, Spark, etc.). This programming framework will, further, allow the hardware accelerators to be swapped in and out of the heterogeneous infrastructure so as to offer high flexibility and energy efficiency. VINEYARD will foster the expansion of the soft-IP core industry, currently limited in the embedded systems, to the data-centre market. VINEYARD plans to demonstrate the advantages of its approach in three real use-cases (a) a bio-informatics application for high-accuracy brain modeling, (b) two critical financial applications, and (c) a big-data analysis application.

Emerging applications such as generative artificial intelligence (AI) and deep neural networks are extremely computationally intensive, making traditional CPU-based infrastructure insufficient for handling complexity and scale. Accelerators like GPUs and field-programmable gate arrays have become essential to efficiently process the vast datasets and complex models that these applications require. Hyperscale data centers deploy thousands of accelerators to meet workload demand; however, as more enterprises recognize the strategic importance of owning and customizing their generative AI models, many are shifting toward deploying accelerator-based data centers on their premises. This shift presents unique opportunities, including improved data security, latency reduction, and the flexibility to customize models in-house. However, it also introduces significant challenges, including high initial capital expenditure, the need for specialized infrastructure and cooling solutions, and the complexity of managing these advanced resources. IT professionals need to analyze several factors (technological, economic, and operational) carefully before deploying accelerators in their premises.