diff --git a/vpr/src/place/net_cost_handler.cpp b/vpr/src/place/net_cost_handler.cpp
index 65ec74dbb47..9f0c23fccd4 100644
--- a/vpr/src/place/net_cost_handler.cpp
+++ b/vpr/src/place/net_cost_handler.cpp
@@ -84,14 +84,6 @@ static void add_block_to_bb(const t_physical_tile_loc& new_pin_loc,
                             t_2D_bb& bb_edge_new,
                             t_2D_bb& bb_coord_new);
 
-/**
- * @brief Given the 3D BB, calculate the wire-length estimate of the net
- * @param net_id ID of the net which wirelength estimate is requested
- * @param bb Bounding box of the net
- * @return Wirelength estimate of the net
- */
-static double get_net_wirelength_estimate(ClusterNetId net_id, const t_bb& bb);
-
 /**
  * @brief To get the wirelength cost/est, BB perimeter is multiplied by a factor to approximately correct for the half-perimeter
  * bounding box wirelength's underestimate of wiring for nets with fanout greater than 2.
@@ -275,7 +267,7 @@ std::pair<double, double> NetCostHandler::comp_cube_bb_cost_(e_cost_methods meth
             net_cost_[net_id] = get_net_cube_bb_cost_(net_id, /*use_ts=*/false);
             cost += net_cost_[net_id];
             if (method == e_cost_methods::CHECK) {
-                expected_wirelength += get_net_wirelength_estimate(net_id, place_move_ctx.bb_coords[net_id]);
+                expected_wirelength += get_net_wirelength_estimate_(net_id);
             }
         }
     }
@@ -1419,7 +1411,9 @@ double NetCostHandler::get_net_per_layer_bb_cost_(ClusterNetId net_id, bool use_
     return ncost;
 }
 
-static double get_net_wirelength_estimate(ClusterNetId net_id, const t_bb& bb) {
+double NetCostHandler::get_net_wirelength_estimate_(ClusterNetId net_id) const {
+    const auto& move_ctx = placer_state_.move();
+    const t_bb& bb = move_ctx.bb_coords[net_id];
     auto& cluster_ctx = g_vpr_ctx.clustering();
 
     double crossing = wirelength_crossing_count(cluster_ctx.clb_nlist.net_pins(net_id).size());
@@ -1438,23 +1432,27 @@ static double get_net_wirelength_estimate(ClusterNetId net_id, const t_bb& bb) {
     return ncost;
 }
 
-double NetCostHandler::get_net_wirelength_from_layer_bb_(ClusterNetId net_id) {
+double NetCostHandler::get_net_wirelength_from_layer_bb_(ClusterNetId net_id) const {
     /* WMF: Finds the estimate of wirelength due to one net by looking at   *
      * its coordinate bounding box.                                         */
 
     const auto& move_ctx = placer_state_.move();
     const std::vector<t_2D_bb>& bb = move_ctx.layer_bb_coords[net_id];
-    const auto& layer_pin_sink_count = move_ctx.num_sink_pin_layer[size_t(net_id)];
+    const vtr::NdMatrixProxy<int, 1> net_layer_pin_sink_count = move_ctx.num_sink_pin_layer[size_t(net_id)];
 
     double ncost = 0.;
-    const int num_layers = g_vpr_ctx.device().grid.get_num_layers();
+    VTR_ASSERT_SAFE(static_cast<int>(bb.size()) == g_vpr_ctx.device().grid.get_num_layers());
 
-    for (int layer_num = 0; layer_num < num_layers; layer_num++) {
-        VTR_ASSERT_SAFE(layer_pin_sink_count[layer_num] != OPEN);
-        if (layer_pin_sink_count[layer_num] == 0) {
+    for (size_t layer_num = 0; layer_num < bb.size(); layer_num++) {
+        VTR_ASSERT_SAFE(net_layer_pin_sink_count[layer_num] != OPEN);
+        if (net_layer_pin_sink_count[layer_num] == 0) {
             continue;
         }
-        double crossing = wirelength_crossing_count(layer_pin_sink_count[layer_num] + 1);
+
+        // The reason we add 1 to the number of sink pins is because when per-layer bounding box is used,
+        // we want to get the estimated wirelength of the given layer assuming that the source pin is
+        // also on that layer
+        double crossing = wirelength_crossing_count(net_layer_pin_sink_count[layer_num] + 1);
 
         /* Could insert a check for xmin == xmax.  In that case, assume  *
          * connection will be made with no bends and hence no x-cost.    *
@@ -1627,12 +1625,15 @@ void NetCostHandler::recompute_costs_from_scratch(const PlaceDelayModel* delay_m
 
 double NetCostHandler::get_total_wirelength_estimate() const {
     const auto& clb_nlist = g_vpr_ctx.clustering().clb_nlist;
-    const auto& bb_coords = placer_state_.move().bb_coords;
 
     double estimated_wirelength = 0.0;
     for (ClusterNetId net_id : clb_nlist.nets()) { /* for each net ... */
         if (!clb_nlist.net_is_ignored(net_id)) {   /* Do only if not ignored. */
-            estimated_wirelength += get_net_wirelength_estimate(net_id, bb_coords[net_id]);
+            if (cube_bb_) {
+                estimated_wirelength += get_net_wirelength_estimate_(net_id);
+            } else {
+                estimated_wirelength += get_net_wirelength_from_layer_bb_(net_id);
+            }
         }
     }
 
diff --git a/vpr/src/place/net_cost_handler.h b/vpr/src/place/net_cost_handler.h
index 510ffa60653..99bda916575 100644
--- a/vpr/src/place/net_cost_handler.h
+++ b/vpr/src/place/net_cost_handler.h
@@ -508,14 +508,6 @@ class NetCostHandler {
      */
     double get_net_per_layer_bb_cost_(ClusterNetId net_id, bool use_ts);
 
-    /**
-     * @brief Given the per-layer BB, calculate the wire-length estimate of the net on each layer
-     * and return the sum of the lengths
-     * @param net_id ID of the net which wirelength estimate is requested
-     * @return Wirelength estimate of the net
-     */
-    double get_net_wirelength_from_layer_bb_(ClusterNetId net_id);
-
     /**
      * @brief Computes the inverse of average channel width for horizontal and
      * vertical channels within a bounding box.
@@ -548,4 +540,21 @@ class NetCostHandler {
      * @return ChanZ cost factor
      */
     float get_chanz_cost_factor_(const t_bb& bb);
+
+    /**
+     * @brief Given the 3D BB, calculate the wire-length estimate of the net
+     * @param net_id ID of the net which wirelength estimate is requested
+     * @param bb Bounding box of the net
+     * @return Wirelength estimate of the net
+     */
+    double get_net_wirelength_estimate_(ClusterNetId net_id) const;
+
+    /**
+     * @brief Given the per-layer BB, calculate the wire-length estimate of the net on each layer
+     * and return the sum of the lengths
+     * @param bb Per-layer BB of the net
+     * @param net_layer_pin_sink_count Number of sink pins on each layer for the net
+     * @return Wirelength estimate of the net
+     */
+    double get_net_wirelength_from_layer_bb_(ClusterNetId net_id) const;
 };
diff --git a/vtr_flow/tasks/regression_tests/vtr_reg_strong/strong_3d/config/config.txt b/vtr_flow/tasks/regression_tests/vtr_reg_strong/strong_3d/config/config.txt
new file mode 100644
index 00000000000..245fb466a3d
--- /dev/null
+++ b/vtr_flow/tasks/regression_tests/vtr_reg_strong/strong_3d/config/config.txt
@@ -0,0 +1,35 @@
+#
+############################################
+# Configuration file for running experiments
+##############################################
+
+# Path to directory of circuits to use
+circuits_dir=benchmarks/titan_blif/other_benchmarks/stratixiv
+
+# Path to directory of SDC files to use
+sdc_dir=benchmarks/titan_blif/other_benchmarks/stratixiv
+
+# Path to directory of architectures to use
+archs_dir=arch/multi_die/stratixiv_3d
+
+# Add circuits to list to sweep
+circuit_list_add=ucsb_152_tap_fir_stratixiv_arch_timing.blif
+
+# Add architectures to list to sweep
+arch_list_add=3d_SB_inter_die_stratixiv_arch.timing.xml
+arch_list_add=3d_full_OPIN_inter_die_stratixiv_arch.timing.xml
+
+# Parse info and how to parse
+parse_file=vpr_titan.txt
+
+# How to parse QoR info
+qor_parse_file=qor_vpr_titan.txt
+
+# Pass requirements
+pass_requirements_file=pass_requirements_vpr_titan.txt
+
+script_params=-starting_stage vpr -track_memory_usage --route_chan_width 300 --max_router_iterations 400 --router_lookahead map
+script_params_list_add = --place_bounding_box_mode auto_bb
+script_params_list_add = --place_bounding_box_mode cube_bb
+script_params_list_add = --place_bounding_box_mode per_layer_bb
+
diff --git a/vtr_flow/tasks/regression_tests/vtr_reg_strong/strong_3d/config/golden_results.txt b/vtr_flow/tasks/regression_tests/vtr_reg_strong/strong_3d/config/golden_results.txt
new file mode 100644
index 00000000000..22cc148a91e
--- /dev/null
+++ b/vtr_flow/tasks/regression_tests/vtr_reg_strong/strong_3d/config/golden_results.txt
@@ -0,0 +1,7 @@
+arch	circuit	script_params	vtr_flow_elapsed_time	vtr_max_mem_stage	vtr_max_mem	error	num_io	num_LAB	num_DSP	num_M9K	num_M144K	num_PLL	vpr_status	vpr_revision	vpr_build_info	vpr_compiler	vpr_compiled	hostname	rundir	max_vpr_mem	num_primary_inputs	num_primary_outputs	num_pre_packed_nets	num_pre_packed_blocks	num_netlist_clocks	num_post_packed_nets	num_post_packed_blocks	device_width	device_height	device_grid_tiles	device_limiting_resources	device_name	pack_mem	pack_time	initial_placed_wirelength_est	placed_wirelength_est	total_swap	accepted_swap	rejected_swap	aborted_swap	place_mem	place_time	place_quench_time	initial_placed_CPD_est	placed_CPD_est	placed_setup_TNS_est	placed_setup_WNS_est	placed_geomean_nonvirtual_intradomain_critical_path_delay_est	place_delay_matrix_lookup_time	place_quench_timing_analysis_time	place_quench_sta_time	place_total_timing_analysis_time	place_total_sta_time	routed_wirelength	avg_routed_wirelength	routed_wiresegment	avg_routed_wiresegment	total_nets_routed	total_connections_routed	total_heap_pushes	total_heap_pops	logic_block_area_total	logic_block_area_used	routing_area_total	routing_area_per_tile	crit_path_route_success_iteration	num_rr_graph_nodes	num_rr_graph_edges	collapsed_nodes	critical_path_delay	geomean_nonvirtual_intradomain_critical_path_delay	setup_TNS	setup_WNS	hold_TNS	hold_WNS	create_rr_graph_time	create_intra_cluster_rr_graph_time	adding_internal_edges	route_mem	crit_path_route_time	crit_path_total_timing_analysis_time	crit_path_total_sta_time	router_lookahead_mem	tile_lookahead_computation_time	router_lookahead_computation_time	
+3d_SB_inter_die_stratixiv_arch.timing.xml	ucsb_152_tap_fir_stratixiv_arch_timing.blif	common_--place_bounding_box_mode_auto_bb	83.45	vpr	1.36 GiB		42	758	0	0	0	0	success	v8.0.0-12389-g509012469-dirty	Release IPO VTR_ASSERT_LEVEL=2	GNU 10.3.0 on Linux-4.15.0-213-generic x86_64	2025-04-15T09:49:51	betzgrp-wintermute.eecg.utoronto.ca	/home/mohagh18/vtr-verilog-to-routing/vtr_flow/tasks	1421664	13	29	26295	20086	1	12439	800	29	21	1218	LAB	auto	1073.4 MiB	12.25	180137	58272	230944	40790	173771	16383	1388.3 MiB	10.48	0.15	5.04678	4.86539	-4206.86	-3.86539	2.46284	0.05	0.0444933	0.0387879	3.15031	2.6241	99517	8.00169	32308	2.59773	27919	38196	46358210	11164094	0	0	2.54084e+07	20860.8	14	2001132	6214436	-1	4.98283	2.70637	-5461.41	-3.98283	0	0	12.21	-1	-1	1388.3 MiB	10.28	4.91142	4.17713	1388.3 MiB	-1	16.75	
+3d_SB_inter_die_stratixiv_arch.timing.xml	ucsb_152_tap_fir_stratixiv_arch_timing.blif	common_--place_bounding_box_mode_cube_bb	85.51	vpr	1.36 GiB		42	758	0	0	0	0	success	v8.0.0-12389-g509012469-dirty	Release IPO VTR_ASSERT_LEVEL=2	GNU 10.3.0 on Linux-4.15.0-213-generic x86_64	2025-04-15T09:49:51	betzgrp-wintermute.eecg.utoronto.ca	/home/mohagh18/vtr-verilog-to-routing/vtr_flow/tasks	1421424	13	29	26295	20086	1	12439	800	29	21	1218	LAB	auto	1073.5 MiB	12.33	180137	58272	230944	40790	173771	16383	1388.1 MiB	10.48	0.14	5.04678	4.86539	-4206.86	-3.86539	2.46284	0.05	0.0434223	0.0377288	3.13504	2.6098	99517	8.00169	32308	2.59773	27919	38196	46358210	11164094	0	0	2.54084e+07	20860.8	14	2001132	6214436	-1	4.98283	2.70637	-5461.41	-3.98283	0	0	13.53	-1	-1	1388.1 MiB	10.72	4.93468	4.19984	1388.1 MiB	-1	17.21	
+3d_SB_inter_die_stratixiv_arch.timing.xml	ucsb_152_tap_fir_stratixiv_arch_timing.blif	common_--place_bounding_box_mode_per_layer_bb	86.05	vpr	1.36 GiB		42	758	0	0	0	0	success	v8.0.0-12389-g509012469-dirty	Release IPO VTR_ASSERT_LEVEL=2	GNU 10.3.0 on Linux-4.15.0-213-generic x86_64	2025-04-15T09:49:51	betzgrp-wintermute.eecg.utoronto.ca	/home/mohagh18/vtr-verilog-to-routing/vtr_flow/tasks	1421572	13	29	26295	20086	1	12439	800	29	21	1218	LAB	auto	1073.7 MiB	12.02	186170	63595	242080	43083	181450	17547	1388.3 MiB	12.05	0.17	5.04678	4.86192	-4242.28	-3.86192	2.41884	0.05	0.051916	0.0457829	3.35985	2.75261	103428	8.31615	32795	2.63689	27768	38066	44034475	9785894	0	0	2.54084e+07	20860.8	14	2001132	6214436	-1	5.18643	2.65983	-5392.13	-4.18642	0	0	13.06	-1	-1	1388.3 MiB	9.86	5.14157	4.32891	1388.3 MiB	-1	18.69	
+3d_full_OPIN_inter_die_stratixiv_arch.timing.xml	ucsb_152_tap_fir_stratixiv_arch_timing.blif	common_--place_bounding_box_mode_auto_bb	57.03	vpr	1.19 GiB		42	758	0	0	0	0	success	v8.0.0-12389-g509012469-dirty	Release IPO VTR_ASSERT_LEVEL=2	GNU 10.3.0 on Linux-4.15.0-213-generic x86_64	2025-04-15T09:49:51	betzgrp-wintermute.eecg.utoronto.ca	/home/mohagh18/vtr-verilog-to-routing/vtr_flow/tasks	1245688	13	29	26295	20086	1	12439	800	29	21	1218	LAB	auto	1073.9 MiB	11.85	186170	63157	219808	34278	166444	19086	1216.5 MiB	10.07	0.18	5.41016	4.9834	-5385.92	-3.9834	3.13071	0.01	0.0545904	0.0489007	2.99902	2.51676	73601	5.91791	18330	1.47383	25639	35167	11163573	1688400	0	0	2.60031e+07	21349.0	15	354380	4692432	-1	5.20923	2.79354	-5293.11	-4.20923	0	0	9.46	-1	-1	1216.5 MiB	4.58	4.95133	4.23051	1216.5 MiB	-1	1.96	
+3d_full_OPIN_inter_die_stratixiv_arch.timing.xml	ucsb_152_tap_fir_stratixiv_arch_timing.blif	common_--place_bounding_box_mode_cube_bb	56.04	vpr	1.19 GiB		42	758	0	0	0	0	success	v8.0.0-12389-g509012469-dirty	Release IPO VTR_ASSERT_LEVEL=2	GNU 10.3.0 on Linux-4.15.0-213-generic x86_64	2025-04-15T09:49:51	betzgrp-wintermute.eecg.utoronto.ca	/home/mohagh18/vtr-verilog-to-routing/vtr_flow/tasks	1245528	13	29	26295	20086	1	12439	800	29	21	1218	LAB	auto	1073.7 MiB	11.88	180137	61714	223520	36703	169137	17680	1216.3 MiB	10.63	0.16	5.41016	4.96403	-5546.44	-3.96403	2.84288	0.01	0.0440263	0.0385965	3.36781	2.8222	75346	6.05821	18897	1.51942	26061	36573	12725206	1711712	0	0	2.60031e+07	21349.0	14	354380	4692432	-1	5.08769	2.56235	-5100.1	-4.08769	0	0	7.77	-1	-1	1216.3 MiB	4.84	5.1781	4.41964	1216.3 MiB	-1	1.92	
+3d_full_OPIN_inter_die_stratixiv_arch.timing.xml	ucsb_152_tap_fir_stratixiv_arch_timing.blif	common_--place_bounding_box_mode_per_layer_bb	58.79	vpr	1.19 GiB		42	758	0	0	0	0	success	v8.0.0-12389-g509012469-dirty	Release IPO VTR_ASSERT_LEVEL=2	GNU 10.3.0 on Linux-4.15.0-213-generic x86_64	2025-04-15T09:49:51	betzgrp-wintermute.eecg.utoronto.ca	/home/mohagh18/vtr-verilog-to-routing/vtr_flow/tasks	1245324	13	29	26295	20086	1	12439	800	29	21	1218	LAB	auto	1073.5 MiB	11.84	186170	63157	219808	34278	166444	19086	1216.1 MiB	10.25	0.14	5.41016	4.9834	-5385.92	-3.9834	3.13071	0.01	0.0422071	0.0368357	2.98723	2.48631	73601	5.91791	18330	1.47383	25639	35167	11163573	1688400	0	0	2.60031e+07	21349.0	15	354380	4692432	-1	5.20923	2.79354	-5293.11	-4.20923	0	0	9.89	-1	-1	1216.1 MiB	4.79	4.83395	4.11932	1216.1 MiB	-1	1.80	
diff --git a/vtr_flow/tasks/regression_tests/vtr_reg_strong/task_list.txt b/vtr_flow/tasks/regression_tests/vtr_reg_strong/task_list.txt
index dc70228462c..ab6f93e0e56 100644
--- a/vtr_flow/tasks/regression_tests/vtr_reg_strong/task_list.txt
+++ b/vtr_flow/tasks/regression_tests/vtr_reg_strong/task_list.txt
@@ -102,3 +102,4 @@ regression_tests/vtr_reg_strong/strong_timing_no_fail
 regression_tests/vtr_reg_strong/strong_noc
 regression_tests/vtr_reg_strong/strong_flat_router
 regression_tests/vtr_reg_strong/strong_routing_constraints
+regression_tests/vtr_reg_strong/strong_3d