diff --git a/src/runtime/Makefile b/src/runtime/Makefile
index 91d3eb976c..88ef8077ff 100644
--- a/src/runtime/Makefile
+++ b/src/runtime/Makefile
@@ -511,6 +511,8 @@ ifneq (,$(QEMUCMD))
 
     DEFENABLEVCPUPINNING_NV = true
 
+    DEFENABLENUMA_NV = true
+
     # NVIDIA profile: rootfs filesystem type (erofs for read-only, compressed images)
     DEFROOTFSTYPE_NV := $(ROOTFSTYPE_EROFS)
 
@@ -689,6 +691,7 @@ USER_VARS += DEFAULTTIMEOUT_NV
 USER_VARS += DEFAULTLAUNCHPROCESSTIMEOUT_NV
 USER_VARS += DEFSANDBOXCGROUPONLY_NV
 USER_VARS += DEFENABLEVCPUPINNING_NV
+USER_VARS += DEFENABLENUMA_NV
 USER_VARS += DEFROOTFSTYPE_NV
 USER_VARS += DEFROOTFSTYPE
 USER_VARS += MACHINETYPE
diff --git a/src/runtime/config/configuration-qemu-nvidia-gpu-snp.toml.in b/src/runtime/config/configuration-qemu-nvidia-gpu-snp.toml.in
index 4dae978b9b..b15186867d 100644
--- a/src/runtime/config/configuration-qemu-nvidia-gpu-snp.toml.in
+++ b/src/runtime/config/configuration-qemu-nvidia-gpu-snp.toml.in
@@ -360,7 +360,12 @@ enable_iommu_platform = false
 # as is: map VM NUMA nodes to host 1:1 and bind vCPUs to related CPUs.
 # Note: To take proper advantage of NUMA, static_sandbox_resource_mgmt should
 # also be enabled for memory pre-allocation.
-enable_numa = false
+#
+# GPU workloads strongly benefit from NUMA awareness: when enabled, the runtime
+# validates that each cold-plugged VFIO device (GPU) resides on a host NUMA
+# node covered by the guest NUMA topology, ensuring memory locality. Consider
+# enabling this on multi-NUMA hosts with GPU passthrough.
+enable_numa = @DEFENABLENUMA_NV@
 
 # NUMA node mapping allows customizing how VM NUMA nodes map to host NUMA nodes.
 # Each entry defines a VM NUMA node and the host NUMA node(s) it maps to.
diff --git a/src/runtime/config/configuration-qemu-nvidia-gpu-tdx.toml.in b/src/runtime/config/configuration-qemu-nvidia-gpu-tdx.toml.in
index 1c1ce20b01..2928389b1c 100644
--- a/src/runtime/config/configuration-qemu-nvidia-gpu-tdx.toml.in
+++ b/src/runtime/config/configuration-qemu-nvidia-gpu-tdx.toml.in
@@ -337,7 +337,12 @@ enable_iommu_platform = false
 # as is: map VM NUMA nodes to host 1:1 and bind vCPUs to related CPUs.
 # Note: To take proper advantage of NUMA, static_sandbox_resource_mgmt should
 # also be enabled for memory pre-allocation.
-enable_numa = false
+#
+# GPU workloads strongly benefit from NUMA awareness: when enabled, the runtime
+# validates that each cold-plugged VFIO device (GPU) resides on a host NUMA
+# node covered by the guest NUMA topology, ensuring memory locality. Consider
+# enabling this on multi-NUMA hosts with GPU passthrough.
+enable_numa = @DEFENABLENUMA_NV@
 
 # NUMA node mapping allows customizing how VM NUMA nodes map to host NUMA nodes.
 # Each entry defines a VM NUMA node and the host NUMA node(s) it maps to.
diff --git a/src/runtime/config/configuration-qemu-nvidia-gpu.toml.in b/src/runtime/config/configuration-qemu-nvidia-gpu.toml.in
index 49f9db0d6e..f373082129 100644
--- a/src/runtime/config/configuration-qemu-nvidia-gpu.toml.in
+++ b/src/runtime/config/configuration-qemu-nvidia-gpu.toml.in
@@ -319,7 +319,12 @@ enable_iommu_platform = false
 # as is: map VM NUMA nodes to host 1:1 and bind vCPUs to related CPUs.
 # Note: To take proper advantage of NUMA, static_sandbox_resource_mgmt should
 # also be enabled for memory pre-allocation.
-enable_numa = false
+#
+# GPU workloads strongly benefit from NUMA awareness: when enabled, the runtime
+# validates that each cold-plugged VFIO device (GPU) resides on a host NUMA
+# node covered by the guest NUMA topology, ensuring memory locality. Consider
+# enabling this on multi-NUMA hosts with GPU passthrough.
+enable_numa = @DEFENABLENUMA_NV@
 
 # NUMA node mapping allows customizing how VM NUMA nodes map to host NUMA nodes.
 # Each entry defines a VM NUMA node and the host NUMA node(s) it maps to.
diff --git a/src/runtime/config/configuration-qemu.toml.in b/src/runtime/config/configuration-qemu.toml.in
index 811884a088..5a51f628ca 100644
--- a/src/runtime/config/configuration-qemu.toml.in
+++ b/src/runtime/config/configuration-qemu.toml.in
@@ -314,6 +314,11 @@ enable_iommu_platform = false
 # as is: map VM NUMA nodes to host 1:1 and bind vCPUs to related CPUs.
 # Note: To take proper advantage of NUMA, static_sandbox_resource_mgmt should
 # also be enabled for memory pre-allocation.
+#
+# When VFIO devices (e.g. GPUs) are cold-plugged and NUMA is enabled, the
+# runtime validates that each device's host NUMA node is covered by the guest
+# NUMA topology. A warning is logged if a device falls outside the configured
+# nodes, indicating potential cross-NUMA memory access overhead.
 enable_numa = false
 
 # NUMA node mapping allows customizing how VM NUMA nodes map to host NUMA nodes.