minimun dimensions of tensors in cudnn library

In cudnn user guide v6.0, page 31. Tensors are restricted to having at least 4 dimensions, Is that true? Because in the official example code RNN, the tensors are 3 dimensions, as

dimA[0] = miniBatch;
dimA[1] = inputSize;
dimA[2] = 1;

strideA[0] = dimA[2] * dimA[1];
strideA[1] = dimA[2];
strideA[2] = 1;

cudnnErrCheck(cudnnSetTensorNdDescriptor(xDesc[i], CUDNN_DATA_FLOAT, 3, dimA, strideA));
cudnnErrCheck(cudnnSetTensorNdDescriptor(dxDesc[i], CUDNN_DATA_FLOAT, 3, dimA, strideA));

Can anyone explain this? Thank you very much.