diff --git a/projects/clr/rocclr/runtime/device/pal/paldevice.cpp b/projects/clr/rocclr/runtime/device/pal/paldevice.cpp
index a03c69f90b..2b1f78a470 100644
--- a/projects/clr/rocclr/runtime/device/pal/paldevice.cpp
+++ b/projects/clr/rocclr/runtime/device/pal/paldevice.cpp
@@ -731,6 +731,8 @@ Device::create(Pal::IDevice* device)
     numDmaEngines_ =
         properties().engineProperties[Pal::EngineTypeDma].engineCount;
 
+    // Creates device settings
+    settings_ = new pal::Settings();
     Pal::PalPublicSettings*const palSettings = iDev()->GetPublicSettings();
     // Modify settings here
     // palSettings ...
@@ -740,6 +742,7 @@ Device::create(Pal::IDevice* device)
     palSettings->cmdBufBatchedSubmitChainLimit = 0;
     palSettings->disableResourceProcessingManager = true;
     palSettings->disableScManager = true;
+    palSettings->numScratchWavesPerCu = settings().numScratchWavesPerCu_;
 
     // Commit the new settings for the device
     result = iDev()->CommitSettingsAndInit();
@@ -768,8 +771,6 @@ Device::create(Pal::IDevice* device)
     Pal::WorkStationCaps wscaps = {};
     iDev()->QueryWorkStationCaps(&wscaps);
 
-    // Creates device settings
-    settings_ = new pal::Settings();
     pal::Settings* gpuSettings = reinterpret_cast<pal::Settings*>(settings_);
     if ((gpuSettings == nullptr) || !gpuSettings->create(properties(), heaps,
          wscaps, appProfile_.reportAsOCL12Device())) {
@@ -1202,7 +1203,6 @@ Device::createBuffer(
             LogError("Can't get the owner object for subbuffer allocation");
             return nullptr;
         }
-
         return gpuParent->createBufferView(owner);
     }
 
@@ -1838,7 +1838,7 @@ Device::allocScratch(uint regNum, const VirtualGPU* vgpu)
                     // Calculate the size of the scratch buffer for a queue
                     uint32_t numTotalCUs = info().maxComputeUnits_;
                     uint32_t numMaxWaves =
-                        properties().gfxipProperties.shaderCore.maxScratchWavesPerCu * numTotalCUs;
+                        settings().numScratchWavesPerCu_* numTotalCUs;
                     scratchBuf->size_ = static_cast<uint64_t>(properties().
                         gfxipProperties.shaderCore.wavefrontSize) *
                         scratchBuf->regNum_ * numMaxWaves * sizeof(uint32_t);
diff --git a/projects/clr/rocclr/runtime/device/pal/palsettings.cpp b/projects/clr/rocclr/runtime/device/pal/palsettings.cpp
index 7aa4080150..9a90c4f91e 100644
--- a/projects/clr/rocclr/runtime/device/pal/palsettings.cpp
+++ b/projects/clr/rocclr/runtime/device/pal/palsettings.cpp
@@ -124,6 +124,8 @@ Settings::Settings()
     numDeviceEvents_ = 1024;
     numWaitEvents_   = 8;
 
+    numScratchWavesPerCu_ = 16;
+
     // Don't support platform atomics by default.
     svmAtomics_ = false;
 
diff --git a/projects/clr/rocclr/runtime/device/pal/palsettings.hpp b/projects/clr/rocclr/runtime/device/pal/palsettings.hpp
index c8e33b94fb..0f9f23e648 100644
--- a/projects/clr/rocclr/runtime/device/pal/palsettings.hpp
+++ b/projects/clr/rocclr/runtime/device/pal/palsettings.hpp
@@ -88,6 +88,7 @@ public:
     uint    numDeviceEvents_;   //!< The number of device events
     uint    numWaitEvents_;     //!< The number of wait events for device enqueue
     uint    hostMemDirectAccess_;   //!< Enables direct access to the host memory
+    uint    numScratchWavesPerCu_;  //!< Maximum number of waves when scratch is enabled
     size_t  xferBufSize_;       //!< Transfer buffer size for image copy optimization
     size_t  stagedXferSize_;    //!< Staged buffer size
     size_t  pinnedXferSize_;    //!< Pinned buffer size for transfer