diff --git a/include/hsakmttypes.h b/include/hsakmttypes.h
index 5d6df38f95..5ac51c2208 100644
--- a/include/hsakmttypes.h
+++ b/include/hsakmttypes.h
@@ -533,7 +533,8 @@ typedef struct _HsaMemFlags
 					    // The KFD will ensure that the memory returned is allocated in the optimal memory location
 					    // and optimal alignment requirements
             unsigned int FixedAddress : 1; // Allocate memory at specified virtual address. Fail if address is not free.
-            unsigned int Reserved    : 16;
+            unsigned int NoNUMABind:    1; // Don't bind system memory to a specific NUMA node
+            unsigned int Reserved    : 15;
 
         } ui32;
         HSAuint32 Value;
diff --git a/src/fmm.c b/src/fmm.c
index e7c83d3fd3..a3f6e7b6a2 100644
--- a/src/fmm.c
+++ b/src/fmm.c
@@ -1397,7 +1397,7 @@ void *fmm_allocate_doorbell(uint32_t gpu_id, uint64_t MemorySizeInBytes,
 		flags.Value = 0;
 		flags.ui32.NonPaged = 1;
 		flags.ui32.HostAccess = 1;
-		flags.ui32.Reserved = 0xBe11;
+		flags.ui32.Reserved = 0xBe1;
 
 		pthread_mutex_lock(&aperture->fmm_mutex);
 		vm_obj->flags = flags.Value;
@@ -1462,10 +1462,13 @@ static int bind_mem_to_numa(uint32_t node_id, void *mem,
 	int num_node;
 	long r;
 
+	if (flags.ui32.NoNUMABind)
+		return 0;
+
 	if (numa_available() == -1)
 		return 0;
 
-	num_node = numa_num_task_nodes();
+	num_node = numa_max_node();
 
 	/* Ignore binding requests to invalid nodes IDs */
 	if (node_id >= (unsigned)num_node) {
diff --git a/src/queues.c b/src/queues.c
index e0cca33c58..8505bdbb1a 100644
--- a/src/queues.c
+++ b/src/queues.c
@@ -424,6 +424,7 @@ void *allocate_exec_aligned_memory_gpu(uint32_t size, uint32_t align,
 	flags.ui32.NonPaged = nonPaged;
 	flags.ui32.PageSize = HSA_PAGE_SIZE_4KB;
 	flags.ui32.CoarseGrain = DeviceLocal;
+	flags.ui32.NoNUMABind = 1;
 
 	size = ALIGN_UP(size, align);