diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 387dd8901..2202b39ef 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -1812,7 +1812,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     // AMD GCN graphics cards perform best when the number of rows per shader is doubled
     uint32_t rm = 1;
-    if ((device->vendor_id == VK_VENDOR_ID_AMD) && (device->subgroup_size == 64))
+    if ((device->subgroup_size_control) && (device->vendor_id == VK_VENDOR_ID_AMD) && (device->subgroup_min_size == 64) && (device->subgroup_max_size == 64))
         rm = 2;
 
     // computing additional rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.