diff --git a/globals.cpp b/globals.cpp
index 25fabb5ead..312f8d0d59 100644
--- a/globals.cpp
+++ b/globals.cpp
@@ -49,5 +49,7 @@ int vendor_packet_process;
 int enable_vendor_packet;
 /* check available system memory before allocation */
 bool check_avail_sysram = false;
+/* enable 128MB sub-allocator in libhsakmt, default: disable */
+int enable_thunk_sub_allocator = 0;
 
-size_t max_single_alloc_size = 0;
\ No newline at end of file
+size_t max_single_alloc_size = 0;
diff --git a/libhsakmt.h b/libhsakmt.h
index 392db59afa..c6c4ffe565 100644
--- a/libhsakmt.h
+++ b/libhsakmt.h
@@ -48,6 +48,7 @@ extern int vendor_packet_process;
 extern int enable_vendor_packet;
 extern bool check_avail_sysram;
 extern size_t max_single_alloc_size;
+extern int enable_thunk_sub_allocator;
 
 #undef HSAKMTAPI
 #define HSAKMTAPI __attribute__((visibility ("default")))
diff --git a/memory.cpp b/memory.cpp
index 2b65def00b..57b7710fc7 100644
--- a/memory.cpp
+++ b/memory.cpp
@@ -278,7 +278,8 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtAllocMemoryAlign(HSAuint32 PreferredNode,
                                                void **MemoryAddress) {
   return hsaKmtAllocMemoryAlignInternal(PreferredNode, SizeInBytes,
                                         Alignment, MemFlags,
-                                        MemoryAddress);
+                                        MemoryAddress,
+                                        !enable_thunk_sub_allocator);
 }
 
 HSAKMT_STATUS hsaKmtFreeMemoryInternal(void *MemoryAddress,
diff --git a/openclose.cpp b/openclose.cpp
index 55f48c1e7e..6bf3ded56f 100644
--- a/openclose.cpp
+++ b/openclose.cpp
@@ -132,6 +132,10 @@ static HSAKMT_STATUS init_vars_from_env(void) {
   if (envvar)
     check_avail_sysram = !strcmp(envvar, "1");
 
+  envvar = getenv("WSL_ENABLE_THUNK_SUB_ALLOCATOR");
+  if (envvar)
+    enable_thunk_sub_allocator = atoi(envvar);
+
   return HSAKMT_STATUS_SUCCESS;
 }