diff --git a/src/database/database.cpp b/src/database/database.cpp
index 60851fe7..fca3102d 100644
--- a/src/database/database.cpp
+++ b/src/database/database.cpp
@@ -99,7 +99,8 @@ Database::Database(const Device &device, const std::string &kernel_name,
     if (device.Type() == "CPU") {
       const auto extensions = device.Capabilities();
       const auto is_apple = (extensions.find("cl_APPLE_SetMemObjectDestructor") == std::string::npos) ? false : true;
-      if (is_apple) {
+      const auto is_likely_apple = device.MaxWorkGroupSize() <= 32;
+      if (is_apple || is_likely_apple) {
         databases.push_front(apple_cpu_fallback);
       }
     }
diff --git a/src/routines/common.cpp b/src/routines/common.cpp
index e81c7873..d3c402bd 100644
--- a/src/routines/common.cpp
+++ b/src/routines/common.cpp
@@ -38,7 +38,8 @@ void RunKernel(Kernel &kernel, Queue &queue, const Device &device,
     auto local_size = size_t{1};
     for (auto &item: local) { local_size *= item; }
     if (local_size > device.MaxWorkGroupSize()) {
-      throw RuntimeErrorCode(StatusCode::kInvalidLocalThreadsTotal);
+      throw RuntimeErrorCode(StatusCode::kInvalidLocalThreadsTotal,
+                             ToString(local_size) + " is larger than " + ToString(device.MaxWorkGroupSize()));
     }
 
     // Make sure the global thread sizes are at least equal to the local sizes