diff --git a/paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc
index 114daaecb59369658191b382a0471d30448a7462..78862a5559dab78dbce2af48c7b10febed39287a 100644
--- a/paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc
@@ -144,6 +144,7 @@ class ConcatMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         platform::errors::InvalidArgument(
             "The axis is expected to be in range of [%d, %d), but got %d",
             -rank, rank, concat_axis));
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     if (concat_axis < 0) {
       concat_axis = concat_axis + rank;
     }
diff --git a/paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc
index 6f0987deeabf50de1cd91a5f7fc0a461b35fa1f6..9e0bdeee6b38b127e96fdbe0015a228ac720750b 100644
--- a/paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc
@@ -561,6 +561,7 @@ class FCMKLDNNOpKernel : public framework::OpKernel<T_in> {
     PADDLE_ENFORCE_EQ(
         platform::is_cpu_place(ctx.GetPlace()), true,
         platform::errors::PreconditionNotMet("FC MKL-DNN must use CPUPlace."));
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     auto input = ctx.Input<LoDTensor>("Input");
     auto w = ctx.Input<Tensor>("W");
     auto bias = ctx.Input<Tensor>("Bias");
diff --git a/paddle/fluid/operators/mkldnn/matmul_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/matmul_mkldnn_op.cc
index 1f2216cbed2b256b15df21956da6741affd8b296..9f1fcf5bd0fbeb49a114c37c6c39bd1612a3359e 100644
--- a/paddle/fluid/operators/mkldnn/matmul_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/matmul_mkldnn_op.cc
@@ -379,6 +379,7 @@ class DNNLMatMulKernel : public framework::OpKernel<T> {
                         platform::errors::Unimplemented(
                             "DNNL matmul doesn't support multiple heads."));
     }
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     ExecuteMatMul<T, T>(ctx);
   }
 };
diff --git a/paddle/fluid/operators/mkldnn/mul_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/mul_mkldnn_op.cc
index 4f0b7cab47efe33fbb297542975590b44f62c75a..5abb7bf406a979bd9aedf4bf7e7d713b82dbba69 100644
--- a/paddle/fluid/operators/mkldnn/mul_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/mul_mkldnn_op.cc
@@ -342,6 +342,7 @@ class MulMKLDNNKernel : public framework::OpKernel<XT> {
     PADDLE_ENFORCE_EQ(platform::is_cpu_place(ctx.GetPlace()), true,
                       paddle::platform::errors::PreconditionNotMet(
                           "Operator DNNL Mul must use CPUPlace"));
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
     const auto &mkldnn_engine = dev_ctx.GetEngine();
 
diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc
index 4922fbeacc619eaa6b3e189ff81b292a9e035ea0..297466e8e5a624359406c5551941ceaa73e5c5c5 100644
--- a/paddle/fluid/platform/device_context.cc
+++ b/paddle/fluid/platform/device_context.cc
@@ -466,6 +466,15 @@ MKLDNNDeviceContextThreadLocals::Body::get_cur_paddle_data_layout(void) {
   return cur_paddle_data_layout;
 }
 
+void MKLDNNDeviceContextThreadLocals::Body::log_lib_version(void) {
+  if (!said_once) {
+    said_once = true;
+    auto dv = dnnl::version();
+    LOG(INFO) << "oneDNN v" << dv->major << "." << dv->minor << "."
+              << dv->patch;
+  }
+}
+
 void MKLDNNDeviceContext::ResetBlobMap() {
   std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
   if (!block_next_cache_clearing_) {
diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h
index 620e2d41c13aff8c70b0b96487cf8d541699d64b..de4c4a8363552d4ddc61de31052c997fc76a39c8 100644
--- a/paddle/fluid/platform/device_context.h
+++ b/paddle/fluid/platform/device_context.h
@@ -466,6 +466,7 @@ class MKLDNNDeviceContextThreadLocals {
 
   typedef MKLDNNDeviceContextThreadLocals self;
   struct Body {
+    bool said_once = false;
     size_t cur_mkldnn_session_id;
     // Current data input shape string.
     // - For fixed-shape, it's a null string in default.
@@ -485,6 +486,7 @@ class MKLDNNDeviceContextThreadLocals {
     void set_cur_input_shape_cache_capacity(int input_shape_cache_capacity);
     void set_cur_paddle_data_layout(framework::DataLayout dl);
     framework::DataLayout get_cur_paddle_data_layout(void);
+    void log_lib_version(void);
   };
   MKLDNNDeviceContextThreadLocals() = default;
   MKLDNNDeviceContextThreadLocals(const MKLDNNDeviceContextThreadLocals& c) =
diff --git a/paddle/fluid/platform/mkldnn_reuse.h b/paddle/fluid/platform/mkldnn_reuse.h
index 90266f6c2099b9667a78f9ca6d29c7ceec2a74bb..96f8fd29c7964c0a21156413f27dcafea9a1eaea 100644
--- a/paddle/fluid/platform/mkldnn_reuse.h
+++ b/paddle/fluid/platform/mkldnn_reuse.h
@@ -45,6 +45,7 @@ class MKLDNNHandlerT {
         key_common_(base_key),
         fwd_pd_(nullptr),
         bwd_pd_(nullptr) {
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     if (platform::MKLDNNDeviceContext::tls().get_cur_mkldnn_session_id() !=
         platform::MKLDNNDeviceContextThreadLocals::kMKLDNNSessionID_Default) {
       key_ = key_common_;
@@ -311,6 +312,7 @@ class MKLDNNHandler {
   MKLDNNHandler(const MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
                 const std::string& base_key)
       : dev_ctx_(dev_ctx), engine_(engine), key_common_(base_key) {
+    platform::MKLDNNDeviceContext::tls().log_lib_version();
     if (platform::MKLDNNDeviceContext::tls().get_cur_mkldnn_session_id() !=
         platform::MKLDNNDeviceContextThreadLocals::kMKLDNNSessionID_Default) {
       key_ = key_common_;