diff --git a/level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl b/level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
index eca5110721..64928bd3bb 100644
--- a/level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
+++ b/level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
@@ -362,6 +362,12 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
 
     auto maxWgCountPerTile = kernel->getMaxWgCountPerTile(this->engineGroupType);
 
+    auto isFlushL3ForExternalAllocationRequired = isFlushL3AfterPostSync && isKernelUsingExternalAllocation;
+    auto isFlushL3ForHostUsmRequired = isFlushL3AfterPostSync && isKernelUsingSystemAllocation;
+    if (NEO::debugManager.flags.DisableFlushL3ForHostUsm.get() && isFlushL3ForHostUsmRequired) {
+        isFlushL3ForExternalAllocationRequired = true;
+        isFlushL3ForHostUsmRequired = false;
+    }
     NEO::EncodeKernelArgsExt dispatchKernelArgsExt = {};
 
     NEO::EncodeDispatchKernelArgs dispatchKernelArgs{
@@ -390,8 +396,8 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
             .isUsingSystemAllocation = isKernelUsingSystemAllocation,
             .dcFlushEnable = this->dcFlushSupport,
             .interruptEvent = interruptEvent,
-            .isFlushL3ForExternalAllocationRequired = isFlushL3AfterPostSync && isKernelUsingExternalAllocation,
-            .isFlushL3ForHostUsmRequired = isFlushL3AfterPostSync && isKernelUsingSystemAllocation,
+            .isFlushL3ForExternalAllocationRequired = isFlushL3ForExternalAllocationRequired,
+            .isFlushL3ForHostUsmRequired = isFlushL3ForHostUsmRequired,
         },
         .preemptionMode = kernelPreemptionMode,
         .requiredPartitionDim = launchParams.requiredPartitionDim,
diff --git a/opencl/source/command_queue/hardware_interface_xehp_and_later.inl b/opencl/source/command_queue/hardware_interface_xehp_and_later.inl
index 3e656c7fd3..8fa989d4d4 100644
--- a/opencl/source/command_queue/hardware_interface_xehp_and_later.inl
+++ b/opencl/source/command_queue/hardware_interface_xehp_and_later.inl
@@ -107,6 +107,11 @@ inline void HardwareInterface<GfxFamily>::programWalker(
             bool flushL3AfterPostSyncForHostUsm = kernelSystemAllocation;
             bool flushL3AfterPostSyncForExternalAllocation = kernel.isUsingSharedObjArgs();
 
+            if (debugManager.flags.DisableFlushL3ForHostUsm.get() && flushL3AfterPostSyncForHostUsm) {
+                flushL3AfterPostSyncForHostUsm = false;
+                flushL3AfterPostSyncForExternalAllocation = true;
+            }
+
             GpgpuWalkerHelper<GfxFamily>::template setupTimestampPacketFlushL3<WalkerType>(&walkerCmd, productHelper, flushL3AfterPostSyncForHostUsm, flushL3AfterPostSyncForExternalAllocation);
         }
     }
diff --git a/shared/source/debug_settings/debug_variables_base.inl b/shared/source/debug_settings/debug_variables_base.inl
index aa61a2772a..517c09b7ad 100644
--- a/shared/source/debug_settings/debug_variables_base.inl
+++ b/shared/source/debug_settings/debug_variables_base.inl
@@ -627,6 +627,7 @@ DECLARE_DEBUG_VARIABLE(bool, EnableBOChunkingPrefetch, false, "Enables prefetchi
 DECLARE_DEBUG_VARIABLE(bool, EnableBOChunkingDevMemPrefetch, false, "Enables prefetching of Device Memory chunks")
 DECLARE_DEBUG_VARIABLE(bool, EnableBOChunkingPreferredLocationHint, false, "Enables preferred location advise on chunks")
 DECLARE_DEBUG_VARIABLE(bool, EnableCompatibilityMode, true, "Enables compatibility mode for platforms which can use precompiled base platform configuration")
+DECLARE_DEBUG_VARIABLE(bool, DisableFlushL3ForHostUsm, false, "Disables L3 flush for host usm")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableBOChunking, -1, "Enables use of chunking of BOs in the KMD, mask: -1 = default, 0 = no chunking, 1 = shared allocations only, 2 = multi-tile device allocations only, 3 = shared and multi-tile device allocations .")
 DECLARE_DEBUG_VARIABLE(int32_t, DestroyAllocationsViaGmm, -1, "Use DeAllocate2 wrapper instead of raw GDI destroy allocations")
 DECLARE_DEBUG_VARIABLE(int32_t, NumberOfBOChunks, 2, "Number of chunks to use")
diff --git a/shared/test/common/test_files/igdrcl.config b/shared/test/common/test_files/igdrcl.config
index 92c0343f1a..02a58ee0a8 100644
--- a/shared/test/common/test_files/igdrcl.config
+++ b/shared/test/common/test_files/igdrcl.config
@@ -545,6 +545,7 @@ EnableBOChunkingDevMemPrefetch = 0
 EnableBOChunkingPreferredLocationHint = 0
 DestroyAllocationsViaGmm = -1
 EnableCompatibilityMode = 1
+DisableFlushL3ForHostUsm = 0
 NumberOfBOChunks = 2
 SetBOChunkingSize = -1
 EnableBOChunking = -1