Reducing the Cost of GPU Cold Starts in Serverless Deep Learning Inference Serving | IEEE Conference Publication | IEEE Xplore