diff --git a/inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp b/inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp
index aef93029ce1..61e817c854c 100644
--- a/inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp
+++ b/inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp
@@ -187,12 +187,20 @@ void prepare_primitive_fusing::fuse_reorders(program &p) {
 void prepare_primitive_fusing::fuse_activations(program &p) {
     bool is_debug = p.get_options().get<build_option_type::debug>()->enabled();
     std::map<primitive_id, std::vector<primitive_id>> fusing_history;
+    bool use_onednn_impls = false;
+
+#ifdef ENABLE_ONEDNN_FOR_GPU
+    auto& engine = p.get_engine();
+    if (engine.get_device_info().supports_immad && engine.configuration().queue_type == queue_types::in_order)
+        use_onednn_impls = true;
+#endif
+
     auto itr = p.get_processing_order().begin();
     while (itr != p.get_processing_order().end()) {
         auto node_itr = itr++;
         auto& node = (*node_itr);
 
-        program_helpers::do_for_types<activation>(*node, [&p, &is_debug, &fusing_history](activation_node& node) {
+        program_helpers::do_for_types<activation>(*node, [&p, &is_debug, &fusing_history, &use_onednn_impls](activation_node& node) {
             auto& input = node.input();
             auto id = node.id();
             // Restrictions:
@@ -239,6 +247,9 @@ void prepare_primitive_fusing::fuse_activations(program &p) {
                     return;
             }
 
+            if (input.is_type<reshape>() && use_onednn_impls)
+                return;
+
             if (input.get_fused_primitives().empty()) {
                 input.add_fused_activation(node.get_primitive()->activation_function, node.get_primitive()->additional_params);
                 for (size_t i = 0; i < node.get_fused_activations_funcs().size(); i++) {
diff --git a/inference-engine/thirdparty/clDNN/src/graph_optimizer/remove_redundant_reorders.cpp b/inference-engine/thirdparty/clDNN/src/graph_optimizer/remove_redundant_reorders.cpp
index 2737572aa0a..07a9f63a9a0 100644
--- a/inference-engine/thirdparty/clDNN/src/graph_optimizer/remove_redundant_reorders.cpp
+++ b/inference-engine/thirdparty/clDNN/src/graph_optimizer/remove_redundant_reorders.cpp
@@ -56,7 +56,12 @@ void remove_redundant_reorders::run(program& p) {
 
             // Avoid different data types between input and output
             auto same_data_type = input.get_output_layout().data_type == output_layout.data_type;
-            if (!same_data_type)
+            auto i8_u8_input = input.get_output_layout().data_type == data_types::i8 ||
+                               input.get_output_layout().data_type == data_types::u8;
+            auto quantize_user = node.get_users().front()->is_type<quantize>() &&
+                                 node.get_users().size() == 1;
+
+            if (!same_data_type && !(i8_u8_input && quantize_user))
                 continue;
 
             // Avoid optimization of nv12 reorder
@@ -323,7 +328,7 @@ void remove_redundant_reorders::run(program& p) {
         auto& dep = node_ptr->get_dependency(0);
         if (!usr->is_type<quantize>() ||
             (dep.get_output_layout().format != format::b_fs_yx_fsv16 &&
-             dep.get_output_layout().format != format::fs_b_yx_fsv32 &&
+             (lo.get_optimization_attributes().use_onednn_impls || dep.get_output_layout().format != format::fs_b_yx_fsv32) &&
              dep.get_output_layout().format != format::bfyx))
             continue;
 
@@ -348,7 +353,7 @@ void remove_redundant_reorders::run(program& p) {
         auto& dep = node->get_dependency(0);
 
         if (!(usr->is_type<convolution>()) ||
-            usr->get_output_layout().data_type != dep.get_output_layout().data_type ||
+            node->get_output_layout().data_type != dep.get_output_layout().data_type ||
             dep.get_output_layout().format != format::bfyx)
             return false;
         if (usr->as<convolution>().get_preferred_impl_type() != impl_types::onednn &&
diff --git a/inference-engine/thirdparty/clDNN/src/layout_optimizer.cpp b/inference-engine/thirdparty/clDNN/src/layout_optimizer.cpp
index 2bf19225e76..34f3c1498cf 100644
--- a/inference-engine/thirdparty/clDNN/src/layout_optimizer.cpp
+++ b/inference-engine/thirdparty/clDNN/src/layout_optimizer.cpp
@@ -178,6 +178,8 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
     auto prev_output_layout = prev.get_output_layout();
     auto next_output_layout = next.get_output_layout();
     auto prev_dt = prev.get_output_layout().data_type;
+    auto next_dt = next.get_output_layout().data_type;
+    auto use_onednn_impls = _optimization_attributes.use_onednn_impls;
 
     auto is_input_idx = [&](size_t idx) -> bool {
         if (&next.get_dependency(idx) == &prev)
@@ -191,7 +193,7 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
         return true;
 
     if (next.is_type<pooling>() &&
-        ((prev_simple && next_simple) ||
+        (((prev_simple && next_simple) && (prev_dt == next_dt)) ||
         ((fmt_prev == format::b_fs_yx_fsv4 && fmt_next == format::bfyx) && (prev_dt == data_types::u8 || prev_dt == data_types::i8))))
         return true;
 
@@ -223,9 +225,12 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
 
     // Additional check: fmt_prev == fmt_next is added only when onednn is enabled.
     if (next.is_type<convolution>() &&
-        fmt_prev == format::bfyx && (!get_optimization_attributes().use_onednn_impls || fmt_prev == fmt_next) &&
+        (!use_onednn_impls || fmt_prev == fmt_next) &&
+        (fmt_prev == format::bfyx || fmt_prev == format::bs_fs_yx_bsv4_fsv2) &&
         ((fmt_next == format::fs_b_yx_fsv32 && next.as<convolution>().get_primitive()->groups == 1) ||
         (fmt_next == format::b_fs_yx_fsv32 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
+        (fmt_next == format::bs_fs_yx_bsv32_fsv32 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
+        (fmt_next == format::bs_fs_yx_bsv32_fsv16 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
         (fmt_next == format::bs_fs_yx_bsv16_fsv16 && next_output_layout.size.feature[0] % 16 == 0 && prev_output_layout.size.feature[0] == 3) ||
         (fmt_next == format::bs_fs_yx_bsv16_fsv16 && next_output_layout.size.feature[0] >= 16 && prev_output_layout.size.feature[0] == 3 &&
         (next_output_layout.data_type != data_types::i8 && next_output_layout.data_type != data_types::u8))))
@@ -239,8 +244,10 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
         return true;
 
     if (next.is_type<convolution>() &&
-        fmt_prev == format::b_fs_yx_fsv4 &&
+        (fmt_prev == format::b_fs_yx_fsv4 || fmt_prev == format::bs_fs_yx_bsv4_fsv4) &&
         ((fmt_next == format::b_fs_yx_fsv32 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
+        (fmt_next == format::bs_fs_yx_bsv32_fsv32 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
+        (fmt_next == format::bs_fs_yx_bsv4_fsv4 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
         (fmt_next == format::b_fs_yx_fsv16 && next_output_layout.size.feature[0] >= 16 &&
         (prev_output_layout.size.feature[0] == 3 || (prev_output_layout.size.feature[0] == 4 && (prev_dt == data_types::u8 || prev_dt == data_types::i8))))))
         return true;
@@ -263,12 +270,51 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
             next_output_layout.size.feature[0] >= 16 && prev_output_layout.size.feature[0] == 3))
         return true;
 
+    if (use_onednn_impls) {
+        if (next.is_type<eltwise>() && (fmt_prev == format::bfyx) && (fmt_next == format::bs_fs_yx_bsv4_fsv2) &&
+            prev.is_input() && (prev_dt == data_types::u8 || prev_dt == data_types::i8))
+            return true;
+
+        if (next.is_type<convolution>() && (prev.is_type<eltwise>() || prev.is_type<quantize>()) &&
+            (fmt_prev == format::bfyx || fmt_prev == format::bs_fs_yx_bsv4_fsv2) &&
+            ((fmt_next == format::bs_fs_yx_bsv32_fsv32 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
+            (fmt_next == format::bs_fs_yx_bsv32_fsv16 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4))))
+            return true;
+
+        if (next.is_type<quantize>())
+            return true;
+
+        if (next.is_type<permute>()) {
+            auto is_rotating_except_batch = [](const std::vector<uint16_t>& order) {
+                // Target transform: Rotate feature dim to back to be taken as inner-most axis
+                // ex) 0(b), 4(f), 1(z), 2(y), 3(x)
+                // ex) 0(b), 3(f), 1(y), 2(x)
+                if ((int32_t) order[1] != order.size() - 1) return false;
+                if ((int32_t) order[0] != 0) return false;
+                for (int32_t i = 2; i < (int32_t) order.size(); ++i) {
+                    if ((int32_t)order[i] !=  (i - 1)) return false;
+                }
+                return true;
+            };
+
+            auto& permute_order = next.as<permute>().get_primitive()->permute_order;
+            if ((fmt_prev == format::b_fs_yx_fsv4 || fmt_prev == format::b_fs_yx_fsv32 || fmt_prev == format::b_fs_zyx_fsv32 ||
+                fmt_prev == format::b_fs_yx_fsv16 || fmt_prev == format::b_fs_zyx_fsv16 || fmt_prev == format::bs_fs_yx_bsv16_fsv16)
+                && permute_order[1] == 2
+                && (!is_rotating_except_batch(permute_order))) {
+                    return false;
+            }
+            return true;
+        }
+    }
+
     return false;
 }
 
 bool layout_optimizer::can_fuse_reorder_to_prev(program_node& prev, program_node& next, format fmt_prev, format fmt_next) {
     auto dt_prev = prev.get_output_layout().data_type;
     auto dt_next = next.get_output_layout().data_type;
+    auto use_onednn_impls = _optimization_attributes.use_onednn_impls;
 
     if (prev.is_type<reorder>())
         return true;
@@ -283,7 +329,7 @@ bool layout_optimizer::can_fuse_reorder_to_prev(program_node& prev, program_node
         return true;
 
     if (prev.is_type<quantize>() &&
-        (fmt_next == format::b_fs_yx_fsv4 || fmt_next == format::b_fs_yx_fsv32 || fmt_next == format::b_fs_zyx_fsv32 ||
+        (fmt_next == format::b_fs_yx_fsv4 || fmt_next == format::b_fs_zyx_fsv32 || (fmt_next == format::b_fs_yx_fsv32 && !use_onednn_impls) ||
          fmt_next == format::b_fs_yx_fsv16 || fmt_next == format::b_fs_zyx_fsv16 || fmt_next == format::bs_fs_yx_bsv16_fsv16))
         return true;
 
@@ -310,6 +356,11 @@ bool layout_optimizer::can_fuse_reorder_to_prev(program_node& prev, program_node
         return true;
     }
 
+    if (use_onednn_impls) {
+        if (prev.is_type<convolution>() && fmt_next == format::bs_fs_yx_bsv32_fsv16 && fmt_prev == format::bs_fs_yx_bsv4_fsv2)
+            return true;
+    }
+
     return false;
 }
 
@@ -1270,17 +1321,60 @@ format layout_optimizer::get_preferred_format(program_node& node) {
             node.as<detection_output>(),
             layout{ data_types::f32, format::bfyx, tensor{} }).format;
     } else if (node.is_type<quantize>()) {
+        bool use_onednn_impls = _optimization_attributes.use_onednn_impls;
         auto layout = node.get_output_layout();
         if (layout.format.spatial_num() == 2 &&
             (layout.data_type == data_types::i8 || layout.data_type == data_types::u8) &&
             layout.size.batch[0] % 16 == 0) {
-            if (layout.size.feature[0] > 8) {
+            if (use_onednn_impls && layout.size.batch[0] % 32 == 0) {
+                if (node.get_users().size() == 1 && node.get_users().front()->is_type<convolution>()) {
+                    auto& conv = node.get_users().front()->as<convolution>();
+                    auto ws = conv.get_dependency(1).get_output_layout().size;
+                    if (data_type_traits::is_floating_point(conv.get_output_layout().data_type) || ws.spatial[0] != 7 || conv.get_primitive()->groups > 1)
+                        expected = format::bfyx;
+                    else
+                        expected = format::bs_fs_yx_bsv4_fsv4;
+
+                    auto conv_output_layout = conv.get_output_layout();
+                    auto weights_layout = conv.weights(0).get_output_layout();
+                    format expected_conv_fmt = get_expected_layout(conv_output_layout, conv, weights_layout).format;
+                    if (expected == format::bfyx && expected_conv_fmt == format::bs_fs_yx_bsv32_fsv32 &&
+                        layout.size.feature[0] % 32 == 0) {
+                        expected = expected_conv_fmt;
+                    }
+                } else {
+                    expected = format::bfyx;
+                }
+            } else if (layout.size.feature[0] > 8) {
                 expected = format::b_fs_yx_fsv16;
             } else {
-                expected = format::b_fs_yx_fsv4;
+                if (use_onednn_impls && layout.size.batch[0] % 16 == 0 && node.get_users().front()->is_type<convolution>()) {
+                    auto& conv = node.get_users().front()->as<convolution>();
+                    auto ws = conv.get_dependency(1).get_output_layout().size;
+                    if (data_type_traits::is_floating_point(conv.get_output_layout().data_type) || ws.spatial[0] != 7 || conv.get_primitive()->groups > 1)
+                        expected = format::bfyx;
+                    else
+                        expected = format::bs_fs_yx_bsv4_fsv4;
+                } else {
+                    expected = format::b_fs_yx_fsv4;
+                }
             }
         } else if (layout.format.spatial_num() == 3 && (layout.data_type == data_types::i8 || layout.data_type == data_types::u8)) {
             expected = format::b_fs_zyx_fsv16;
+        } else if (use_onednn_impls) {
+            if (node.get_users().size() == 1 && node.get_users().front()->is_type<convolution>() &&
+                (layout.data_type == data_types::i8 || layout.data_type == data_types::u8)) {
+                auto& conv = node.get_users().front()->as<convolution>();
+                auto conv_output_layout = conv.get_output_layout();
+                auto weights_layout = conv.weights(0).get_output_layout();
+                format expected_conv_fmt = get_expected_layout(conv_output_layout, conv, weights_layout).format;
+                if (layout.format.spatial_num() == 2 && layout.size.feature[0] % 32 == 0 && expected_conv_fmt == format::b_fs_yx_fsv32)
+                    expected = expected_conv_fmt;
+                else
+                    expected = format::bfyx;
+            } else {
+                expected = format::bfyx;
+            }
         }
     } else if (node.is_type<reorder>() || node.is_type<input_layout>()) {
         expected = node.get_output_layout().format;