The Intelligent Cloud and Edge (ICE) Group is at the forefront of advancing the artificial intelligence infrastructure and addressing the fundamental system issues of combining Cloud and Edge. Our mission is to provide efficient, user-friendly, cross-platform artificial intelligence training and deployment technologies.
Our group’s cutting-edge research areas include deep learning compilation frameworks, optimization of new hardware accelerators, system design for new types of workloads such as graph neural networks, Mixture-of-Experts (MoE), scientific computing, and software-hardware co-design optimization for new intelligent scenarios such as gaming and multimedia.
We take pride in our contributions to the field, as evidenced by several research achievements published in top academic conferences such as OSDI, NSDI, ATC, etc. Our main achievements have been open-sourced as projects such as (opens in new tab)NNFusion (opens in new tab), Rammer (opens in new tab), Roller (opens in new tab), SparTA (opens in new tab), Antares (opens in new tab), Tutel (opens in new tab), etc., and some of these technologies have been applied to product lines such as Xbox, Bing, and Office.
Our research directions are focused on solving the most pressing challenges in building a cost-effective AI infrastructure by harvesting both the power of Cloud and Edge. We aim to advance the field by
- exploring new frontiers in deep learning compilation frameworks, e.g., tensor compilation, full model optimization, etc.
- designing large model inference systems combining Cloud and Edge
- developing decentralized large model inference and training systems
- exploring the capability of new hardware accelerator architecture, e.g., mesh-based AI accelerators
- software-hardware co-designed system for sparse model computation
- designing graph neural networks (GNN) and Mixture-of-Experts (MoE) systems,
- accelerating AI-based workload like databases, gaming, multimedia, etc.
Join us in our pursuit of pioneering AI system technology that will shape the future of the industry.
智能云端系统组（Intelligent Cloud and Edge）致力于研究云端一体的人工智能基础架构及其关键系统问题，以提供高效、易用、跨平台的人工智能训练和部署技术。小组目前研究方向涵盖深度学习编译框架、新型人工智能加速硬件的优化、面向新型负载（如图神经网络、MoE、科学计算等）的系统设计、新型智能场景（如游戏、多媒体等）的软硬件协同优化等。小组的多项研究成果发表在OSDI、NSDI、ATC等顶级学术会议，其中主要成果均以开源项目的形式对外开放（如NNFusion (opens in new tab)、Rammer (opens in new tab)、Roller (opens in new tab)、Antares (opens in new tab)、Tutel (opens in new tab)等），部分技术也被应用在诸如Xbox、Bing、Office等产品线中。
研究方向：
- 深度学习编译框架和编译技术
- 云端结合的大模型的推理系统设计与优化
- 去中心化的大模型推理和训练系统
- 新型硬件加速加的性能优化
- 面向稀疏模型计算的软硬件协同优化
- 图神经网络和MoE系统的设计与优化
- 基于AI的新型负载（如数据库、游戏、多媒体）的支持与加速等