diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlCommandRecorder.h b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlCommandRecorder.h
index 273f172a21..9becf3ad81 100644
--- a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlCommandRecorder.h
+++ b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlCommandRecorder.h
@@ -56,6 +56,11 @@ namespace Dml
         
         void SetAllocator(std::weak_ptr<BucketizedBufferAllocator> allocator);
 
+        bool HasUnsubmittedWork() override
+        {
+            return m_operationsRecordedInCurrentCommandList || !m_pendingCommandLists.empty();
+        }
+
     private:
 
         std::shared_ptr<CommandQueue> m_queue;
diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionContext.cpp b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionContext.cpp
index 6a776f734c..52cbc63fda 100644
--- a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionContext.cpp
+++ b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionContext.cpp
@@ -132,14 +132,6 @@ namespace Dml
         m_dmlRecorder.GetCommandList().CopyTo(commandList);
     }
 
-    void ExecutionContext::Wait(ID3D12Fence* fence, uint64_t value)
-    {
-        assert(!m_closed);
-        Flush();
-        m_queue->Wait(fence, value);
-        ReleaseCompletedReferences();
-    }
-
     void ExecutionContext::SetCommandRecorder(ICommandRecorder* newRecorder)
     {
         assert(!m_closed);
@@ -162,7 +154,7 @@ namespace Dml
     {
         assert(!m_closed);
 
-        if (!m_currentRecorder)
+        if (!m_currentRecorder || !m_currentRecorder->HasUnsubmittedWork())
         {
             // Nothing to flush
             return;
@@ -171,8 +163,11 @@ namespace Dml
         m_currentRecorder->CloseAndExecute();
         ReleaseCompletedReferences();
 
-        // Just submitted our command list, so we have neither DML or D3D12 work recorded on any of our command lists.
+        // Pre-emptively set the DML command recorder.  It's the only command recorder right now,
+        // and doing this here causes work and allocations resetting the command list to occur at
+        // a point where it's going to be parallelized with GPU work.
         m_currentRecorder = nullptr;
+        SetCommandRecorder(&m_dmlRecorder);
     }
     
     void ExecutionContext::QueueReference(IUnknown* object) 
@@ -203,7 +198,7 @@ namespace Dml
 
         // If something has been recorded into a command list but not submitted yet, it means that the *next* fence
         // value is the one to signal completion.
-        const bool unflushedWorkExists = (m_currentRecorder != nullptr);
+        const bool unflushedWorkExists = (m_currentRecorder != nullptr) && m_currentRecorder->HasUnsubmittedWork();
         if (unflushedWorkExists)
         {
             ++event.fenceValue;
diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionContext.h b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionContext.h
index 523d9d5541..f70ee13572 100644
--- a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionContext.h
+++ b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionContext.h
@@ -68,9 +68,6 @@ namespace Dml
 
         void GetCommandListForRecording(ID3D12GraphicsCommandList** commandList);
 
-        // See ID3D12CommandQueue::Wait
-        void Wait(ID3D12Fence* fence, uint64_t value);
-
         // Forces all queued work to begin executing on the GPU. This method returns immediately and does not wait
         // for the submitted work to complete execution on the GPU.
         void Flush();
diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ICommandRecorder.h b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ICommandRecorder.h
index 28e49c8e6f..b0dad7e128 100644
--- a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ICommandRecorder.h
+++ b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/ICommandRecorder.h
@@ -15,6 +15,8 @@ namespace Dml
         // Forces all queued work to begin executing on the GPU. This method returns immediately and does not wait
         // for the submitted work to complete execution on the GPU.
         virtual void CloseAndExecute() = 0;
+
+        virtual bool HasUnsubmittedWork() = 0;
     };
 
 } // namespace Dml