diff --git a/hw/rtl/VX_cluster.sv b/hw/rtl/VX_cluster.sv
index d537249d..90076673 100644
--- a/hw/rtl/VX_cluster.sv
+++ b/hw/rtl/VX_cluster.sv
@@ -167,6 +167,6 @@ module VX_cluster import VX_gpu_pkg::*; #(
         );
     end
 
-    `BUFFER_BUSY (busy, (| per_socket_busy), (`NUM_SOCKETS > 1));
+    `BUFFER_EX(busy, (| per_socket_busy), 1'b1, (`NUM_SOCKETS > 1));
 
 endmodule
diff --git a/hw/rtl/VX_define.vh b/hw/rtl/VX_define.vh
index 5dcf1f2f..95d206ce 100644
--- a/hw/rtl/VX_define.vh
+++ b/hw/rtl/VX_define.vh
@@ -307,20 +307,20 @@
 
 ///////////////////////////////////////////////////////////////////////////////
 
-`define BUFFER_BUSY(dst, src, enable) \
-    logic __busy; \
-    if (enable) begin \
-        always @(posedge clk) begin \
-            if (reset) begin \
-                __busy <= 1'b0; \
-            end else begin \
-                __busy <= src; \
-            end \
-        end \
-    end else begin \
-        assign __busy = src; \
-    end \
-    assign dst = __busy
+`define BUFFER_EX(dst, src, ena, latency) \
+    VX_pipe_register #( \
+        .DATAW  ($bits(dst)), \
+        .RESETW ($bits(dst)), \
+        .DEPTH  (latency) \
+    ) __``dst ( \
+        .clk      (clk), \
+        .reset    (reset), \
+        .enable   (ena), \
+        .data_in  (src), \
+        .data_out (dst) \
+    )
+
+`define BUFFER(dst, src) `BUFFER_EX(dst, src, 1'b1, 1)
 
 `define POP_COUNT_EX(out, in, model) \
     VX_popcount #( \
diff --git a/hw/rtl/VX_socket.sv b/hw/rtl/VX_socket.sv
index 1e61fdff..139598d9 100644
--- a/hw/rtl/VX_socket.sv
+++ b/hw/rtl/VX_socket.sv
@@ -245,6 +245,6 @@ module VX_socket import VX_gpu_pkg::*; #(
         );
     end
 
-    `BUFFER_BUSY (busy, (| per_core_busy), (`SOCKET_SIZE > 1));
+    `BUFFER_EX(busy, (| per_core_busy), 1'b1, (`SOCKET_SIZE > 1));
     
 endmodule
diff --git a/hw/rtl/Vortex.sv b/hw/rtl/Vortex.sv
index 5bd628d5..594204e4 100644
--- a/hw/rtl/Vortex.sv
+++ b/hw/rtl/Vortex.sv
@@ -166,7 +166,7 @@ module Vortex import VX_gpu_pkg::*; (
         );
     end
 
-    `BUFFER_BUSY (busy, (| per_cluster_busy), (`NUM_CLUSTERS > 1));    
+    `BUFFER_EX(busy, (| per_cluster_busy), 1'b1, (`NUM_CLUSTERS > 1));    
 
 `ifdef PERF_ENABLE
 
diff --git a/hw/rtl/core/VX_schedule.sv b/hw/rtl/core/VX_schedule.sv
index 0ffeafc2..9008f605 100644
--- a/hw/rtl/core/VX_schedule.sv
+++ b/hw/rtl/core/VX_schedule.sv
@@ -353,7 +353,7 @@ module VX_schedule import VX_gpu_pkg::*; #(
         .empty     (no_pending_instr)
     );
 
-    `BUFFER_BUSY (busy, (active_warps != 0 || ~no_pending_instr), 1);
+    `BUFFER_EX(busy, (active_warps != 0 || ~no_pending_instr), 1'b1, 1);
 
     // export CSRs
     assign sched_csr_if.cycles = cycles;