ROCm TF 1.3 resnet batch 512 crash inside MIOpen bwd BN

MIOpen(HIP): miopenStatus_t miopenBatchNormalizationBackward(miopenHandle_t, miopenBatchNormMode_t, const void *, const void *, const void *, const void *, const miopenTensorDescriptor_t, const void *, const miopenTensorDescriptor_t, const void *, const miopenTensorDescriptor_t, void *, const miopenTensorDescriptor_t, const void *, void *, void *, double, const void *, const void *){
bn_mode = 1
xDesc = 512, 16, 32, 32
x = 0x63bdc8a00
dyDesc = 512, 16, 32, 32
dy = 0x61fdc8a00
dxDesc = 512, 16, 32, 32
dx = 0x60fdc8a00
bnScaleBiasDiffDesc = 1, 16, 1, 1
bnScale = 0x608e12e00
resultBnScaleDiff = 0x608f07100
resultBnBiasDiff = 0x608f07900
epsilon = 1e-05
savedMean = 0x608f09800
savedInvVariance = 0x608f09900
}
<<hip-api tid:5.3510 5.3510 hipLaunchKernel 'BatchNormBwdSpatial' gridDim:{16,1024,1} groupDim:{1,1024,1} sharedMem:+0 stream:0.1 @328400701367156
Memory access fault by GPU node-1 on address 0x600000000. Reason: Page not present or supervisor privilege.