diff --git a/hw/rtl/VX_define.vh b/hw/rtl/VX_define.vh
index 812c30f1..100204e2 100644
--- a/hw/rtl/VX_define.vh
+++ b/hw/rtl/VX_define.vh
@@ -201,7 +201,7 @@
 `define L2DRAM_TAG_WIDTH    (`L2_ENABLE ? `L2DRAM_ADDR_WIDTH : (`L2DRAM_ADDR_WIDTH+`CLOG2(`NUM_CORES*2)))
 
 // Snoop request tag bits
-`define L2SNP_TAG_WIDTH     ((`NUM_CLUSTERS > 1) ? `LOG2UP(`L3SNRQ_SIZE) : 1)
+`define L2SNP_TAG_WIDTH     ((`NUM_CLUSTERS > 1) ? `LOG2UP(`L3SNRQ_SIZE) : `L3SNP_TAG_WIDTH)
 
 // Number of Word requests per cycle {1, 2, 4, 8, ...}
 `define L2NUM_REQUESTS      (2*`NUM_CORES)
@@ -221,7 +221,7 @@
 `define L3DRAM_TAG_WIDTH    ((`NUM_CLUSTERS > 1) ? `L3DRAM_ADDR_WIDTH : `L2DRAM_TAG_WIDTH)
 
 // Snoop request tag bits
-`define L3SNP_TAG_WIDTH     1 
+`define L3SNP_TAG_WIDTH     16 
 
 // Number of Word requests per cycle {1, 2, 4, 8, ...}
 `define L3NUM_REQUESTS      `NUM_CLUSTERS
diff --git a/hw/simulate/simulator.cpp b/hw/simulate/simulator.cpp
index 27ecb0d7..61cb6ed7 100644
--- a/hw/simulate/simulator.cpp
+++ b/hw/simulate/simulator.cpp
@@ -181,27 +181,34 @@ void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {
   // align address to LLC block boundaries
   auto aligned_addr_start = mem_addr / GLOBAL_BLOCK_SIZE;
   auto aligned_addr_end = (mem_addr + size + GLOBAL_BLOCK_SIZE - 1) / GLOBAL_BLOCK_SIZE;
-  int outstanding_snp_reqs = 0;
-
+  
   // submit snoop requests for the needed blocks
   vortex_->snp_req_addr  = aligned_addr_start;
+  vortex_->snp_req_tag   = 0;
   vortex_->snp_req_valid = 1;
   vortex_->snp_rsp_ready = 1;
+
+  int pending_snp_reqs = 1;
+
   for (;;) {
-    this->step();
-    if (vortex_->snp_rsp_valid) {
-      assert(outstanding_snp_reqs > 0);
-      --outstanding_snp_reqs;
+    this->step();        
+    if (vortex_->snp_rsp_valid) {      
+      assert(pending_snp_reqs > 0);
+      --pending_snp_reqs;
+      //std::cout << timestamp << ": [sim] snp rsp: tag=" << vortex_->snp_rsp_tag << " pending=" << pending_snp_reqs << std::endl;
     }
-    if (vortex_->snp_req_valid && vortex_->snp_req_ready) {
-      ++outstanding_snp_reqs;
-      vortex_->snp_req_addr += 1;
-      if (vortex_->snp_req_addr >= aligned_addr_end) {
+    if (vortex_->snp_req_valid && vortex_->snp_req_ready) {            
+      if (vortex_->snp_req_addr < aligned_addr_end) {        
+        vortex_->snp_req_addr += 1;
+        vortex_->snp_req_tag  += 1;
+        ++pending_snp_reqs;
+        //std::cout << timestamp << ": [sim] snp req: addr=" << vortex_->snp_req_addr << " tag=" << vortex_->snp_req_tag << " remain=" << (aligned_addr_end - vortex_->snp_req_addr) << std::endl;
+      } else {
         vortex_->snp_req_valid = 0;        
-      }
+      }      
     }
     if (!vortex_->snp_req_valid 
-     && 0 == outstanding_snp_reqs) {
+     && 0 == pending_snp_reqs) {
       break;
     }
   }