diff --git a/projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_aql_queue.cpp b/projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_aql_queue.cpp
index 424ec07c86..90123fce31 100644
--- a/projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_aql_queue.cpp
+++ b/projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_aql_queue.cpp
@@ -313,7 +313,7 @@ AqlQueue::AqlQueue(GpuAgent* agent, size_t req_size_pkts, HSAuint32 node_id, Scr
   MAKE_NAMED_SCOPE_GUARD(PM4IBGuard, [&]() { agent_->system_deallocator()(pm4_ib_buf_); });
 
   // Set initial CU mask
-  SetCUMasking(0, nullptr);
+  if (!core::Runtime::runtime_singleton_->flag().cu_mask_skip_init()) SetCUMasking(0, nullptr);
 
   active_ = true;
 
diff --git a/projects/rocr-runtime/runtime/hsa-runtime/core/util/flag.h b/projects/rocr-runtime/runtime/hsa-runtime/core/util/flag.h
index cd395c667d..e6825f7f8a 100644
--- a/projects/rocr-runtime/runtime/hsa-runtime/core/util/flag.h
+++ b/projects/rocr-runtime/runtime/hsa-runtime/core/util/flag.h
@@ -151,6 +151,9 @@ class Flag {
 
     var = os::GetEnvVar("HSA_ENABLE_DEBUG");
     debug_ = (var == "1") ? true : false;
+
+    var = os::GetEnvVar("HSA_CU_MASK_SKIP_INIT");
+    cu_mask_skip_init_ = (var == "1") ? true : false;
   }
 
   void parse_masks(uint32_t maxGpu, uint32_t maxCU) {
@@ -220,6 +223,8 @@ class Flag {
     return it->second;
   }
 
+  bool cu_mask_skip_init() const { return cu_mask_skip_init_; }
+
  private:
   bool check_flat_scratch_;
   bool enable_vm_fault_message_;
@@ -240,6 +245,7 @@ class Flag {
   bool patch_xgmi_link_weight_;
   bool patch_link_override_;
   bool debug_;
+  bool cu_mask_skip_init_;
 
   SDMA_OVERRIDE enable_sdma_;