diff --git a/rpcs3/Emu/RSX/Host/MM.cpp b/rpcs3/Emu/RSX/Host/MM.cpp
index b5c5fb3494..cf21b6e046 100644
--- a/rpcs3/Emu/RSX/Host/MM.cpp
+++ b/rpcs3/Emu/RSX/Host/MM.cpp
@@ -98,6 +98,12 @@ namespace rsx
 			return;
 		}
 
+		std::lock_guard lock(g_mprotect_queue_lock);
+		if (g_deferred_mprotect_queue.empty())
+		{
+			return;
+		}
+
 		auto& rsxdma = g_fxo->get<rsx::dma_manager>();
 		rsxdma.backend_ctrl(mm_backend_ctrl::cmd_mm_flush, nullptr);
 	}
diff --git a/rpcs3/Emu/RSX/NV47/HW/nv47_sync.hpp b/rpcs3/Emu/RSX/NV47/HW/nv47_sync.hpp
index 674c66470d..8bc069ecaf 100644
--- a/rpcs3/Emu/RSX/NV47/HW/nv47_sync.hpp
+++ b/rpcs3/Emu/RSX/NV47/HW/nv47_sync.hpp
@@ -30,7 +30,6 @@ namespace rsx
 				if constexpr (FlushDMA)
 				{
 					// If the backend handled the request, this call will basically be a NOP
-					rsx::mm_flush_lazy();
 					g_fxo->get<rsx::dma_manager>().sync();
 				}
 
@@ -38,10 +37,15 @@ namespace rsx
 				{
 					// Manually flush the pipeline.
 					// It is possible to stream report writes using the host GPU, but that generates too much submit traffic.
-					rsx::mm_flush();
 					RSX(ctx)->sync();
 				}
 
+				if constexpr (FlushDMA || FlushPipe)
+				{
+					// Kick MM flush without waiting. Technically we should do this before the sync in case of MTRSX, but doing it later improves CPU performance.
+					rsx::mm_flush_lazy();
+				}
+
 				if (handled)
 				{
 					// Backend will handle it, nothing to write.