The 2nd International Workshop on High Performance Chiplet and Interconnnect Architectures (HiPChips)
主题
- Optical and other advanced chiplet interconnect technologies
- Interconnect standards of coherent and non-coherent data sharing protocols (e.g. CXL)
- Disaggregated computing architectures powered by high speed interconnect
- Chiplet architectures for in-memory computing and other emerging technologies
- Software optimization and scheduling with fast inter-chiplet network
- Power evaluation and performance modeling of chiplet architectures
议题
Challenges in AI and ML for Chiplets to address
Speaker:Dharmesh Jani (Infrastructure Partnerships/Ecosystems Lead @ Meta)
通常的人类行为场景
- 认知,认识世界并构建认知模型,需要大量的训练
- 挖掘,在各种类型的数据中寻找目标,需要推理
- 综合创造,创造出新的事物
当前AI应用领域及相关算法应用
当前DSA处理器存在算力瓶颈
内存和互联网络拖累了计算。
- 评
Chiplet技术需要去解决的人工智能和机器学习的挑战。该报告针对人类行为包括认知、挖掘、创造三个场景指出了DSA的运行需求,并指出了其挑战主要在加速器和内存间的缺口、软硬件的协同设计、互联网络中如何选通局部带宽以优化总体带宽
Enable Polymorphic AI Architecture via Composable Chiplet Technologies
Speaker: Weifeng Zhang (Chief Architect & VP of Software Lightelligence)
- 评
通过可模组的chiplet技术构建多模态的AI架构。
Building Heterogeneous Chiplets with AMBA Interconnects
Speaker: Jeff Defilippi ( Senior Director Product Management @ ARM)
当前chiplet生态处于各家为战阶段
片上一致性协议总线CHI
新的CHI协议将支持Multi-chip
新的CHI协议正在开发,扩充的新的特征将使CHI从on-chip扩展到multi-chip。
如图所示为IP的层次架构,协议层使用CHI协议,链路层和物理层使用UCIe或其他IP。
- 评
使用AMBA总线构建异步chiplet架构。该报告揭示了chiplet技术应用的迫切需求,介绍了ARM的CHI总线特征,并基于原先on-chip特征的基础上扩展应用于multi-chip的属性,基于CHI的chiplet互联将推动chiplet行业向半定制平台发展
Next-generation Co-Packaged Optics for Future Disaggregated AI Systems
Speaker: Sajjad Moazeni, Assistant Professor (University of Washington, Seattle)