diff --git a/dnn/src/cuda/conv_bias/opr_impl.cpp b/dnn/src/cuda/conv_bias/opr_impl.cpp
index bf37427858ee1ef6abbb78d890c843d1c80dcf2a..17ea6eb945de3b5a64cf9e90ec9defcf1c2d65db 100644
--- a/dnn/src/cuda/conv_bias/opr_impl.cpp
+++ b/dnn/src/cuda/conv_bias/opr_impl.cpp
@@ -145,9 +145,20 @@ ConvBiasForward::Algorithm* ConvBiasForwardImpl::get_algorithm_heuristic(
     const bool prefer_dnn_chanwise = slow_cudnn_chanwise_impl ||
                                      args.filter_meta.stride[0] != 1 ||
                                      args.filter_meta.stride[1] != 1 || hw_size < 512;
+    //! choose for large kernel cases
+    size_t fh = args.filter_meta.spatial[2], fw = args.filter_meta.spatial[3];
+    size_t hi = src[2], wi = src[3];
+    const bool prefer_dnn_lk_implbmm = hi <= 2 * fh && wi <= 2 * fw;
     //! avoid bad case in cudnn, check dnn chanwise impl first
     if (is_chanwise) {
-        if (prefer_dnn_chanwise) {
+        if (prefer_dnn_lk_implbmm) {
+            if (sm_algo_pack.f16_implicit_bmm[0].is_available_attribute(
+                        args, positive_attr, negative_attr, workspace_limit_in_bytes))
+                return &sm_algo_pack.f16_implicit_bmm[0];
+            if (sm_algo_pack.f32_implicit_bmm[0].is_available_attribute(
+                        args, positive_attr, negative_attr, workspace_limit_in_bytes))
+                return &sm_algo_pack.f32_implicit_bmm[0];
+        } else if (prefer_dnn_chanwise) {
             if (sm_algo_pack.chanwise.is_available_attribute(
                         args, positive_attr, negative_attr, workspace_limit_in_bytes))
                 return &sm_algo_pack.chanwise;
diff --git a/dnn/src/cuda/convolution/opr_impl.cpp b/dnn/src/cuda/convolution/opr_impl.cpp
index c39563ca0b38bbc603cab58e70a3e5557239e193..c1d1784d48c83351f65f1103fcbbdcd82a434fd6 100644
--- a/dnn/src/cuda/convolution/opr_impl.cpp
+++ b/dnn/src/cuda/convolution/opr_impl.cpp
@@ -115,6 +115,19 @@ ConvolutionBackwardDataImpl::Algorithm* ConvolutionBackwardDataImpl::
                 const AlgoAttribute& negative_attr) {
     AlgoBase::SizeArgs args(this, filter, diff, grad);
 
+    //! choose for large kernel cases
+    size_t fh = args.filter_meta.spatial[2], fw = args.filter_meta.spatial[3];
+    size_t ho = diff[2], wo = diff[3];
+    const bool prefer_dnn_lk_implbmm = args.filter_meta.format == Param::Format::NCHW &&
+                                       ho <= 2 * fh && wo <= 2 * fw;
+    if (prefer_dnn_lk_implbmm) {
+        if (sm_algo_pack.implbmm_nchw_hmma.is_available_attribute(
+                    args, positive_attr, negative_attr, workspace_limit_in_bytes))
+            return &sm_algo_pack.implbmm_nchw_hmma[0];
+        if (sm_algo_pack.implbmm_nchw_fma.is_available_attribute(args, positive_attr, negative_attr, workspace_limit_in_bytes)) 
+            return &sm_algo_pack.implbmm_nchw_fma[0];
+    }
+
     if (args.filter_meta.group > 1 &&
         sm_algo_pack.chanwise.is_available_attribute(
                 args, positive_attr, negative_attr, workspace_limit_in_bytes)) {