Scholarlink

<p>Deep Neural Networks (DNNs) are widely deployed on Network-on-Chip (NoC)-based many-core accelerators, where efficient workload mapping is critical to performance and energy efficiency. Most existing approaches assume fully connected DNNs and fixed task granularity, assumptions that break down when network pruning introduces irregular and sparse communication patterns. In such scenarios, neuron grouping plays a key role in shaping communication behavior and directly affects mapping quality. This paper proposes a two-stage neuron grouping and mapping framework for NoC-based DNN accelerators. The first stage adaptively groups neurons into logical tasks to improve communication locality and load balance under both fully connected and pruned DNNs. The second stage maps the grouped tasks onto the underlying NoC using an improved Cuckoo optimization strategy with efficient initialization and controlled perturbation. Experimental results show that, compared with the representative NGM-SA baseline, the proposed framework reduces communication cost by an average of 21.01% across different network structures and pruning levels.. Cycle-accurate NoC simulations further demonstrate consistent improvements in latency, energy consumption, and throughput compared with baseline methods.</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 two-stage optimization framework for neuron grouping and mapping in NoC-based DNN accelerators

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Hao Ding,
  • Zhengxun Wen,
  • Xuesong Wang,
  • Junyan Qian

摘要

Deep Neural Networks (DNNs) are widely deployed on Network-on-Chip (NoC)-based many-core accelerators, where efficient workload mapping is critical to performance and energy efficiency. Most existing approaches assume fully connected DNNs and fixed task granularity, assumptions that break down when network pruning introduces irregular and sparse communication patterns. In such scenarios, neuron grouping plays a key role in shaping communication behavior and directly affects mapping quality. This paper proposes a two-stage neuron grouping and mapping framework for NoC-based DNN accelerators. The first stage adaptively groups neurons into logical tasks to improve communication locality and load balance under both fully connected and pruned DNNs. The second stage maps the grouped tasks onto the underlying NoC using an improved Cuckoo optimization strategy with efficient initialization and controlled perturbation. Experimental results show that, compared with the representative NGM-SA baseline, the proposed framework reduces communication cost by an average of 21.01% across different network structures and pruning levels.. Cycle-accurate NoC simulations further demonstrate consistent improvements in latency, energy consumption, and throughput compared with baseline methods.