diff --git a/docs/faq/env_var.md b/docs/faq/env_var.md
index 92cf4931b04b..e373377ee8de 100644
--- a/docs/faq/env_var.md
+++ b/docs/faq/env_var.md
@@ -196,6 +196,12 @@ When USE_PROFILER is enabled in Makefile or CMake, the following environments ca
   - Flag to enable or disable MKLDNN accelerator. On by default.
   - Only applies to mxnet that has been compiled with MKLDNN (```pip install mxnet-mkl``` or built from source with ```USE_MKLDNN=1```)
 
+* MXNET_ENFORCE_DETERMINISM
+  - Values: 0(false) or 1(true) ```(default=0)```
+  - If set to true, MXNet will only use deterministic algorithms in forward and backward computation.
+  If no such algorithm exists given other constraints, MXNet will error out. This variable affects the choice
+  of CUDNN convolution algorithms. Please see [CUDNN developer guide](https://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html) for more details.
+
 Settings for Minimum Memory Usage
 ---------------------------------
 - Make sure ```min(MXNET_EXEC_NUM_TEMP, MXNET_GPU_WORKER_NTHREADS) = 1```
diff --git a/src/operator/nn/cudnn/cudnn_convolution-inl.h b/src/operator/nn/cudnn/cudnn_convolution-inl.h
index d63d46821edc..3bd6c5a3826b 100644
--- a/src/operator/nn/cudnn/cudnn_convolution-inl.h
+++ b/src/operator/nn/cudnn/cudnn_convolution-inl.h
@@ -889,6 +889,7 @@ class CuDNNConvolutionOp {
                        size_t workspace_byte, CuDNNAlgo<AlgoType> *algo) {
     // Determine the fastest acceptable algo that matches the algo_preference (-1 = any),
     // regardless of mathType.
+    bool enforce_determinism = dmlc::GetEnv("MXNET_ENFORCE_DETERMINISM", false);
     for (decltype(perf_results.size()) i = 0; i != perf_results.size(); ++i) {
       const auto &result = perf_results[i];
       bool algo_is_tensor_core = false;
@@ -896,6 +897,9 @@ class CuDNNConvolutionOp {
         algo_is_tensor_core = result.mathType == CUDNN_TENSOR_OP_MATH;
       #endif
       if (result.status == CUDNN_STATUS_SUCCESS &&
+        #if CUDNN_MAJOR >= 7
+          (!enforce_determinism || result.determinism == cudnnDeterminism_t::CUDNN_DETERMINISTIC) &&
+        #endif
           (param_.cudnn_tune.value() != conv::kLimited || result.memory <= workspace_byte)) {
         algo->Set(result.algo, algo_is_tensor_core);
         return;
diff --git a/src/operator/nn/cudnn/cudnn_deconvolution-inl.h b/src/operator/nn/cudnn/cudnn_deconvolution-inl.h
index c0c56507bbf3..72ba2c95fc6a 100644
--- a/src/operator/nn/cudnn/cudnn_deconvolution-inl.h
+++ b/src/operator/nn/cudnn/cudnn_deconvolution-inl.h
@@ -829,6 +829,7 @@ class CuDNNDeconvolutionOp {
   void AlgoFinalSelect(const std::vector<PerfType> &perf_results, std::string kernel_name,
                        size_t workspace_byte, CuDNNAlgo<AlgoType> *algo) {
     // Determine the fastest acceptable algo regardless of mathType.
+    bool enforce_determinism = dmlc::GetEnv("MXNET_ENFORCE_DETERMINISM", false);
     for (decltype(perf_results.size()) i = 0; i != perf_results.size(); ++i) {
       const auto &result = perf_results[i];
       bool algo_is_tensor_core = false;
@@ -836,6 +837,9 @@ class CuDNNDeconvolutionOp {
         algo_is_tensor_core = result.mathType == CUDNN_TENSOR_OP_MATH;
       #endif
       if (result.status == CUDNN_STATUS_SUCCESS &&
+        #if CUDNN_MAJOR >= 7
+          (!enforce_determinism || result.determinism == cudnnDeterminism_t::CUDNN_DETERMINISTIC) &&
+        #endif
           (param_.cudnn_tune.value() != conv::kLimited || result.memory <= workspace_byte)) {
         algo->Set(result.algo, algo_is_tensor_core);
         return;
diff --git a/src/operator/nn/cudnn/cudnn_pooling-inl.h b/src/operator/nn/cudnn/cudnn_pooling-inl.h
index bc3ee366007c..89fa73ef5471 100644
--- a/src/operator/nn/cudnn/cudnn_pooling-inl.h
+++ b/src/operator/nn/cudnn/cudnn_pooling-inl.h
@@ -48,7 +48,12 @@ class CuDNNPoolingOp {
     param_ = p;
     switch (param_.pool_type) {
       case pool_enum::kMaxPooling:
+      #if CUDNN_MAJOR >= 7
+        mode_ = dmlc::GetEnv("MXNET_ENFORCE_DETERMINISM", false) ?
+          CUDNN_POOLING_MAX_DETERMINISTIC : CUDNN_POOLING_MAX;
+      #else
         mode_ = CUDNN_POOLING_MAX;
+      #endif
         break;
       case pool_enum::kAvgPooling:
         if (param_.count_include_pad.has_value() && !param_.count_include_pad.value()) {