Scholarlink

<p>The proliferation of AI applications across diverse domains has driven the evolution of AI accelerators toward higher performance and energy efficiency. This paper addresses the critical challenge of task scheduling in AI accelerators by introducing a fully hardware-managed scheduling system. Our approach leverages Operator Completion Status Registers (OCSRs) and a novel computation-scheduling instruction set to minimize software overhead and maximize execution parallelism. The co-designed hardware-software solution comprises: (1) a dedicated hardware scheduling unit with a complete instruction pipeline, (2) a compiler that maps operators to scheduling instructions while managing OCSR allocation, and (3) a lightweight runtime for efficient task dispatch. Experimental results demonstrate that our system significantly reduces scheduling latency and improves overall throughput, achieving an average performance gain of approximately 30% across multiple CNN models while maintaining minimal area overhead of only 7.41%. The proposed architecture establishes a new paradigm for high-efficiency AI accelerator design.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A fully hardware-managed scheduling architecture for AI accelerators

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月5日

查看全文
  • Libo Cheng,
  • Liang Yang,
  • Jian Shao,
  • Xinyi Gu,
  • Rong Qian,
  • Xinwei Zhang,
  • Donghao Li,
  • Hongbin Wang,
  • Xiaoqi Xia

摘要

The proliferation of AI applications across diverse domains has driven the evolution of AI accelerators toward higher performance and energy efficiency. This paper addresses the critical challenge of task scheduling in AI accelerators by introducing a fully hardware-managed scheduling system. Our approach leverages Operator Completion Status Registers (OCSRs) and a novel computation-scheduling instruction set to minimize software overhead and maximize execution parallelism. The co-designed hardware-software solution comprises: (1) a dedicated hardware scheduling unit with a complete instruction pipeline, (2) a compiler that maps operators to scheduling instructions while managing OCSR allocation, and (3) a lightweight runtime for efficient task dispatch. Experimental results demonstrate that our system significantly reduces scheduling latency and improves overall throughput, achieving an average performance gain of approximately 30% across multiple CNN models while maintaining minimal area overhead of only 7.41%. The proposed architecture establishes a new paradigm for high-efficiency AI accelerator design.