法学硕士的硬件加速：全面调查和比较 Hardware Acceleration of LLMs: A comprehensive survey and comparison

原始链接: https://arxiv.org/abs/2409.03384

论文《LLM 的硬件加速：综合调查和比较》讨论了大型语言模型 (LLM) 在自然语言处理任务中的使用。 它调查了加速这些模型的 Transformer 网络硬件的各种方法，包括 TensorFlow、PyTorch 和 BERT 等框架。 作者根据技术、处理平台（FPGA、ASIC、内存、GPU）、加速、能源效率和性能 (GOP) 等因素对这些框架进行了定性和定量比较。 然而，由于实现技术的差异，公平地比较这些方案具有挑战性。 为了解决这个问题，作者评估了每种方法应用于相同技术时的性能和能源效率。 他们还在多个 FPGA 芯片上测试了 LLM 的某些部分，以实现公平的比较。 本研究的主要目标是确定法学硕士硬件加速的有效方法，以提高自然语言处理能力。

The paper "Hardware Acceleration of LLMs: A Comprehensive Survey and Comparison" discusses the use of large language models (LLMs) for natural language processing tasks. It surveys various approaches to accelerating transformer network hardware for these models, including frameworks such as TensorFlow, PyTorch, and BERT. The authors provide qualitative and quantitative comparisons of these frameworks based on factors like technology, processing platform (FPGA, ASIC, In-Memory, GPU), speedup, energy efficiency, and performance (GOPs). However, due to differences in implementation technologies, it's challenging to compare the schemes fairly. To address this, the authors estimate the performance and energy efficiency of each approach when applied to the same technology. They also test some parts of the LLMs on multiple FPGA chips to achieve a fair comparison. The main goal of this study is to identify efficient methods for hardware acceleration of LLMs for improved natural language processing capabilities.