大语言模型中通信优化推理方向博士研究员

Are you interested in the communication networks behind modern AI systems? Would you like to combine mathematical modeling with experiments on programmable network hardware to make AI infrastructure faster and more energy-efficient? In this PhD project, you will develop models and orchestration methods for the networks that connect GPUs during large language model inference. Are you our next PhD candidate in high-performance AI networking?

Your research challenge Large language models rely on powerful computing infrastructure to generate responses. When inference is distributed across multiple GPUs, these devices must continuously exchange intermediate results. Network congestion and communication overhead can therefore limit performance and increase energy consumption, even when substantial computing resources are available.

In the project Communication-Optimized Inference in Large Language Models (COIL) , you will investigate how analytical network models can support the design and operation of AI clusters. The project combines network performance modeling, optimization, and experimental research, giving you the opportunity to connect mathematical foundations with practical challenges in AI infrastructure.

You will develop fast, composable analytical models using Network Calculus, a mathematical framework for analyzing communication networks. These models will capture interactions between inference traffic, network architecture, and mechanisms such as flow control and congestion feedback. They will provide bounds on communication latency and buffer requirements, complemented by energy estimates based on measurement-calibrated power models.

Building on these models, you will explore how architectural choices—including network topology, link bandwidth, and parallelization strategies—affect performance and energy consumption. You will develop tools that help cluster designers evaluate promising configurations efficiently.

You will also investigate how the models can guide runtime decisions, such as workload placement, prioritization, routing, or resource reservation. A focused prototype will demonstrate how model-informed decisions can improve the operation of an existing network.

Experimental validation will be an integral part of your research. You will use measurements from AI infrastructure and conduct reproducible experiments on programmable networking testbeds. Technologies of interest include Ethernet-based AI networks, emerging Ultra Ethernet mechanisms, P4-programmable switches, and SmartNICs. Finally, you will evaluate the results from the testbeds on real AI infrastructure in collaboration with our partner SURF .

Through this combination of theory, software development, and experimentation, you will contribute to more responsive and energy-efficient AI infrastructure. The project also offers opportunities to share your findings with networking standardization communities and to release research tools and experimental results openly.

An interdisciplinary and international environment You will conduct your research under the main supervision of Prof. Lisa Maile who leads the Network Modeling and Design (NoMaD) lab. NoMaD is part of the Electronic Systems group in the Department of Electrical Engineering at Eindhoven University of Technology. Your supervision will also include partners from the Advanced Models by Open Research and Engineering (AMOR/e) Lab of Computer Science and the Model-Based Design Lab from Electrical Engineering.

The position is funded through the EuroTech programme, which promotes research collaboration between leading technical universities across Europe. As part of this project, you will collaborate closely with Prof. Georg Carle and his research group on Network Architectures and Services at the Technical University of Munich (TUM) . As part of this collaboration, you will spend six-month in Munich with the group at TUM , where you will gain hands-on experience with programmable network hardware and reproducible experimental infrastructure.