⚡ 本页包含 AI 生成的分析内容，仅供参考
论文提出了一种多模态端到端驾驶处理器，通过稀疏性推理单元和灵活的稀疏-密集异构架构最大化稀疏性利用和核心利用率，并设计了节能的分段聚合网络和长短期记忆单元以减少外部内存访问。该处理器在71.3mJ/帧的能耗下达到10.3fps，相比现有先进驾驶SoC能耗降低218倍。
*Equally Credited Authors (ECAs) 1 Abstract A multi-modal end-to-end driving processor is proposed with 4 features: 1) a sparsity reasoning unit to maximize sparsity exploitation, 2) a flexible sparse-dense heterogeneous architecture with a sparsity-aware adaptive core orchestrator to maximize core utilization, 3) an energy-efficient segmented aggregation network, and 4) a long-/short-term memory unit to minimize external memory access (EMA) of temporal attention. It achieves 10.3fps at 71.3mJ/frame, consuming 218× less energy than a state-of-the-art driving SoC. Recently, multimodal sensor-fusion end-to-end driving (EED) models that integrate a CNN and transformer have demonstrated remarkable achievements. By unifying CNN-based feature extraction with transformer-based spatio-temporal correlation learning across tokens in the Bird’s-Eye-View (BEV) space, these models enable breakthrough improvements in
Jueun Jung*1,2, Sangho Lee*1, Junghyun Yoo2, Ghangmin Yun2, Kyuho Jason Lee2
Ulsan National Institute of Science and Technology, Ulsan, Korea, 2Yonsei University, Seoul, Korea