diff --git a/runtime/hsa-runtime/core/inc/amd_gpu_agent.h b/runtime/hsa-runtime/core/inc/amd_gpu_agent.h
index 868a08dbe8..bbbfeb0398 100644
--- a/runtime/hsa-runtime/core/inc/amd_gpu_agent.h
+++ b/runtime/hsa-runtime/core/inc/amd_gpu_agent.h
@@ -536,8 +536,9 @@ class GpuAgent : public GpuAgentInt {
 
   // @brief AQL queues for cache management and blit compute usage.
   enum QueueEnum {
-    QueueUtility,   // Cache management and device to {host,device} blit compute
-    QueueBlitOnly,  // Host to device blit
+    QueueUtility,     // Cache management and device to {host,device} blit compute
+    QueueBlitOnly,    // Host to device blit
+    QueuePCSampling,  // Dedicated high priority queue for PC Sampling
     QueueCount
   };
 
diff --git a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
index 7b230ba736..8268d9a850 100644
--- a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
+++ b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
@@ -741,18 +741,29 @@ core::Blit* GpuAgent::CreateBlitKernel(core::Queue* queue) {
 
 void GpuAgent::InitDma() {
   // Setup lazy init pointers on queues and blits.
-  auto queue_lambda = [this]() {
-    auto ret = CreateInterceptibleQueue();
-    if (ret == nullptr)
+  auto queue_lambda = [this](HSA_QUEUE_PRIORITY priority = HSA_QUEUE_PRIORITY_NORMAL) {
+    auto queue = CreateInterceptibleQueue();
+    if (queue == nullptr)
       throw AMD::hsa_exception(HSA_STATUS_ERROR_OUT_OF_RESOURCES,
                                "Internal queue creation failed.");
-    return ret;
+
+    if (priority != HSA_QUEUE_PRIORITY_NORMAL)
+      if (queue->SetPriority(priority) != HSA_STATUS_SUCCESS)
+        throw AMD::hsa_exception(HSA_STATUS_ERROR,
+                                "Failed to increase queue priority for PC Sampling");
+    return queue;
   };
+
   // Dedicated compute queue for host-to-device blits.
   queues_[QueueBlitOnly].reset(queue_lambda);
   // Share utility queue with device-to-host blits.
   queues_[QueueUtility].reset(queue_lambda);
 
+  // Dedicated compute queue for PC Sampling CP-DMA commands. We need a dedicated queue that runs at
+  // highest priority because we do not want the CP-DMA commands to be delayed/blocked due to
+  // other dispatches/barriers that could be in the other AQL queues.
+  queues_[QueuePCSampling].reset(queue_lambda(HSA_QUEUE_PRIORITY_MAXIMUM));
+
   // Decide which engine to use for blits.
   auto blit_lambda = [this](bool use_xgmi, lazy_ptr<core::Queue>& queue, bool isHostToDev) {
     Flag::SDMA_OVERRIDE sdma_override = core::Runtime::runtime_singleton_->flag().enable_sdma();
diff --git a/runtime/hsa-runtime/core/runtime/hsa_ext_amd.cpp b/runtime/hsa-runtime/core/runtime/hsa_ext_amd.cpp
index 7b56898520..f855ad339a 100644
--- a/runtime/hsa-runtime/core/runtime/hsa_ext_amd.cpp
+++ b/runtime/hsa-runtime/core/runtime/hsa_ext_amd.cpp
@@ -1072,10 +1072,13 @@ hsa_status_t hsa_amd_queue_set_priority(hsa_queue_t* queue,
   core::Queue* cmd_queue = core::Queue::Convert(queue);
   IS_VALID(cmd_queue);
 
+  // Highest queue priority allowed for HSA user is HSA_QUEUE_PRIORITY_HIGH
+  // HSA_QUEUE_PRIORITY_MAXIMUM is reserved for PC Sampling and can only be allocated internally
+  // in ROCR
   static std::map<hsa_amd_queue_priority_t, HSA_QUEUE_PRIORITY> ext_kmt_priomap = {
       {HSA_AMD_QUEUE_PRIORITY_LOW, HSA_QUEUE_PRIORITY_MINIMUM},
       {HSA_AMD_QUEUE_PRIORITY_NORMAL, HSA_QUEUE_PRIORITY_NORMAL},
-      {HSA_AMD_QUEUE_PRIORITY_HIGH, HSA_QUEUE_PRIORITY_MAXIMUM},
+      {HSA_AMD_QUEUE_PRIORITY_HIGH, HSA_QUEUE_PRIORITY_HIGH},
   };
 
   auto priority_it = ext_kmt_priomap.find(priority);