One　of　the　difficulties　encountered　in　realizing　artificial　neural　network　based　on　VLSI　is　the　choice　of　the　implementation　method　of　activation　function.　At　present,　the　main　approaches　to　solve　this　problem　are　piecewise　nonlinear　approximation　and　bit　level　mapping.　Based　on　hyperbolic　tangent,　the　final　output　error　of　the　two　methods　is　discussed　through　the　hardware　implementation　and　software　analysis　of　the　artificial　neural　network　nodes.　We　found　that　the　nonlinear　approximation　method　has　the　problem　of　large　output　fluctuation,　and　the　amplification　effect　of　the　backpropagation　can　not　be　ignored.　Therefore,　this　paper　proposes　that　the　bit　level　mapping　method　has　more　advantages　in　practical　applications　in　the　implementation　of　high-precision　artificial　neural　nodes.　©　2018　IEEE.