diff --git a/inference-engine/thirdparty/clDNN/api/tensor.hpp b/inference-engine/thirdparty/clDNN/api/tensor.hpp
index 09bb93c4d4b..b8ce09b7de9 100644
--- a/inference-engine/thirdparty/clDNN/api/tensor.hpp
+++ b/inference-engine/thirdparty/clDNN/api/tensor.hpp
@@ -390,6 +390,12 @@ struct format {
         }
         return false;
     }
+    /// @brief Checks if @p format is simple data format
+    static bool is_simple_data_format(type fmt) {
+        return (fmt == yxfb || fmt == byxf ||
+                fmt == bfyx || fmt == fyxb ||
+                fmt == bfzyx || fmt == bfwzyx);
+    }
     /// @brief Checks if @p format is of grouped type
     static bool is_grouped(type fmt) { return group_num(fmt) != 0; }
     /// @brief Checks if @p format is of image type
diff --git a/inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/lstm_dynamic_timeloop_ref.cl b/inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/lstm_dynamic_timeloop_ref.cl
index 5703366176f..444b4e87e02 100644
--- a/inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/lstm_dynamic_timeloop_ref.cl
+++ b/inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/lstm_dynamic_timeloop_ref.cl
@@ -61,7 +61,7 @@ KERNEL(lstm_dynamic_timeloop_ref)(
     #else
     bool use_cell = false;
     #endif //cell_term
-   
+
     for(int timestep = 0; timestep < MAX_SEQUENCE_LENGTH; timestep++)
     {
         //not all workitems will do computations
@@ -155,6 +155,6 @@ KERNEL(lstm_dynamic_timeloop_ref)(
         } //second if(timestep < unroll_timesteps)
 
         //all workitems needs to hit the barrier after writing to global output memory
-        barrier(CLK_GLOBAL_MEM_FENCE);         
+        barrier(CLK_GLOBAL_MEM_FENCE);
     }
 }
diff --git a/inference-engine/thirdparty/clDNN/src/gpu/memory_gpu.cpp b/inference-engine/thirdparty/clDNN/src/gpu/memory_gpu.cpp
index ae7f12eac16..347ab507d4a 100644
--- a/inference-engine/thirdparty/clDNN/src/gpu/memory_gpu.cpp
+++ b/inference-engine/thirdparty/clDNN/src/gpu/memory_gpu.cpp
@@ -31,7 +31,7 @@ gpu_buffer::gpu_buffer(const refcounted_obj_ptr<engine_impl>& engine,
                        bool reset)
     : lockable_gpu_mem(engine), memory_impl(engine, layout, net_id, allocation_type::cl_mem, false),
       _buffer(_context->context(), CL_MEM_READ_WRITE, size()) {
-    if (reset && is_memory_reset_needed(_layout)) zero_buffer();
+    if (reset || is_memory_reset_needed(_layout)) zero_buffer();
 }
 
 gpu_buffer::gpu_buffer(const refcounted_obj_ptr<engine_impl>& engine,
@@ -256,7 +256,7 @@ gpu_usm::gpu_usm(const refcounted_obj_ptr<engine_impl>& engine, const layout& la
             "Unknown unified shared memory type!");
     }
 
-    if (reset && is_memory_reset_needed(_layout)) zero_buffer();
+    if (reset || is_memory_reset_needed(_layout)) zero_buffer();
 }
 
 void* gpu_usm::lock() {
diff --git a/inference-engine/thirdparty/clDNN/src/include/memory_impl.h b/inference-engine/thirdparty/clDNN/src/include/memory_impl.h
index e346f545019..0219319994b 100644
--- a/inference-engine/thirdparty/clDNN/src/include/memory_impl.h
+++ b/inference-engine/thirdparty/clDNN/src/include/memory_impl.h
@@ -48,7 +48,7 @@ struct memory_impl : refcounted_obj<memory_impl> {
         // - To be Weights format (Data memory can be reused by memory_pool, which can lead to errors)
         // - To have zero paddings
         // - To be completely filled with data
-        if (!format::is_weights_format(l.format) || format::is_winograd(l.format) || format::is_image_2d(l.format)) {
+        if ((!format::is_weights_format(l.format) && !format::is_simple_data_format(l.format)) || format::is_winograd(l.format) || format::is_image_2d(l.format)) {
             return true;
         }
 
diff --git a/inference-engine/thirdparty/clDNN/src/primitive_inst.cpp b/inference-engine/thirdparty/clDNN/src/primitive_inst.cpp
index 30ec5be2a5e..611baac3308 100644
--- a/inference-engine/thirdparty/clDNN/src/primitive_inst.cpp
+++ b/inference-engine/thirdparty/clDNN/src/primitive_inst.cpp
@@ -188,7 +188,11 @@ memory_impl::ptr primitive_inst::allocate_output() {
                                       false);
     } else if (_network.is_internal() && _node.is_output() && _node.is_type<generic_layer>() &&
                engine.supports_allocation(allocation_type::usm_device)) {
-        return engine.allocate_memory(layout, allocation_type::usm_device, net_id);
+        return engine.allocate_memory(layout, allocation_type::usm_device, net_id, false);
+    } else if (_network.is_internal() && !_node.is_output() && _node.is_type<input_layout>()) {
+        // Skip memory reset for input_layout primitives, since data will be copied from cldnn::data primitive
+        // or just reuse primitive's memory
+        return engine.allocate_memory(layout, alloc_type, net_id, false);
     } else if (_network.is_internal() || (!_node.can_share_buffer()) || _node.can_be_optimized() || _node.is_output()) {
         return engine.allocate_memory(layout, alloc_type, net_id);
     }
diff --git a/inference-engine/thirdparty/clDNN/src/program.cpp b/inference-engine/thirdparty/clDNN/src/program.cpp
index 7718e655ef7..272c81c45ad 100644
--- a/inference-engine/thirdparty/clDNN/src/program.cpp
+++ b/inference-engine/thirdparty/clDNN/src/program.cpp
@@ -532,7 +532,8 @@ void program_impl::transfer_memory_to_device() {
                 // Allocate and transfer memory
                 auto device_mem = mem.get_engine()->allocate_memory(mem.get_layout(),
                                                                     allocation_type::usm_device,
-                                                                    mem.get_net_id());
+                                                                    mem.get_net_id(),
+                                                                    false);
                 dynamic_cast<gpu::gpu_usm&>(*device_mem).copy_from_other(dynamic_cast<gpu::gpu_usm&>(mem));
                 data_node.attach_memory(*device_mem);
                 const_cast<memory&>(data_node.get_primitive()->mem).reset();