diff --git a/hw/rtl/core/VX_schedule.sv b/hw/rtl/core/VX_schedule.sv
index 36bdf9e0..9a074d31 100644
--- a/hw/rtl/core/VX_schedule.sv
+++ b/hw/rtl/core/VX_schedule.sv
@@ -231,7 +231,13 @@ module VX_schedule import VX_gpu_pkg::*; #(
         `ifdef GBAR_ENABLE
             if (warp_ctl_if.valid && warp_ctl_if.barrier.valid
              && warp_ctl_if.barrier.is_global 
+        `ifdef GBAR_CLUSTER_ENABLE
+             // engage cluster barrier as soon as the barrier count is
+             // fulfilled, instead of requiring all warps to be synchronized
+             && (active_barrier_count[`NW_WIDTH-1:0] == warp_ctl_if.barrier.size_m1[`NW_WIDTH-1:0])) begin
+        `else
              && (curr_barrier_mask_n == active_warps)) begin
+        `endif
                 gbar_req_valid <= 1;
                 gbar_req_id <= warp_ctl_if.barrier.id;
                 gbar_req_size_m1 <= warp_ctl_if.barrier.size_m1[`NC_WIDTH-1:0];
diff --git a/hw/rtl/core/VX_wctl_unit.sv b/hw/rtl/core/VX_wctl_unit.sv
index 88b2f71e..5b1ad834 100644
--- a/hw/rtl/core/VX_wctl_unit.sv
+++ b/hw/rtl/core/VX_wctl_unit.sv
@@ -109,7 +109,13 @@ module VX_wctl_unit import VX_gpu_pkg::*; #(
     assign barrier.valid    = is_bar;
     assign barrier.id       = rs1_data[`NB_WIDTH-1:0];
 `ifdef GBAR_ENABLE
+`ifdef GBAR_CLUSTER_ENABLE
+    // all barriers are cluster-wide barriers, which is implemented by
+    // modifying the global barrier logic
+    assign barrier.is_global = 1'b1;
+`else
     assign barrier.is_global = rs1_data[31];
+`endif
 `else
     assign barrier.is_global = 1'b0;
 `endif