Our paper titled "InferCool: Enhancing AI Inference Cooling through Transparent, Non-Intrusive Task Reassignment" has been accepted to the ACM Symposium on Cloud Computing (SoCC). SoCC is a premium conference in cloud computing and data center technologies and this year it will be held at the Microsoft Campus in Redmond, WA in November.
The paper addressed the cooling system challenges for deep learning inference workloads. Leveraging the emerging support for MIG (Multiple-Instance GPU), the paper proposes a novel middleware for data centers that can significantly reduce the cooling costs by re-arranging deep learning inference tasks in a non-intrusive manner.
This work is a collaboration between Paderborn University and the Huazhong University of Science and Technology.
