update

Source/ComputationNetworkLib/ComputationNode.h

@@ -1897,6 +1897,8 @@ public:
     virtual void RequestMatricesBeforeBackprop(MatrixPool& matrixPool) override
     {
         size_t matrixSize = m_sampleLayout.GetNumElements();
+        if (OperationName() == L"DistributedFullyConnected_v2" || OperationName() == L"DistributedAdditiveFullConnection")
+            matrixSize *= Globals::GetProcessNum();
         RequestMatrixFromPool(m_gradient, matrixPool, matrixSize, HasMBLayout(), /*isWorkSpace*/false, ParentGradientReused() || IsGradientReused());
 
         auto multiOutputNode = dynamic_cast<MultiOutputNode<ElemType>*>(this);