[GitHub] eric-haibin-lin commented on issue #8861: The first GPU cost more memory than the others in version 0.12.1

eric-haibin-lin commented on issue #8861: The first GPU cost more memory than the others in version 0.12.1 URL: https://github.com/apache/incubator-mxnet/issues/8861#issuecomment-352565114 If you fix batch size per GPU(120/4=30), and try different number of GPUs, does the first GPU

[GitHub] eric-haibin-lin commented on issue #8861: The first GPU cost more memory than the others in version 0.12.1

eric-haibin-lin commented on issue #8861: The first GPU cost more memory than the others in version 0.12.1 URL: https://github.com/apache/incubator-mxnet/issues/8861#issuecomment-350397214 I didn't see relevant changes regarding kvstore buffer distributed in kvstore comm between 0.10 and

[GitHub] eric-haibin-lin commented on issue #8861: The first GPU cost more memory than the others in version 0.12.1

eric-haibin-lin commented on issue #8861: The first GPU cost more memory than the others in version 0.12.1 URL: https://github.com/apache/incubator-mxnet/issues/8861#issuecomment-348015232 Will take a look. @fanlu what's an easy way to reproduce the problem?