diff --git a/cub/device/dispatch/dispatch_select_if.cuh b/cub/device/dispatch/dispatch_select_if.cuh
index e7a4f1aa8..7d331d3a5 100644
--- a/cub/device/dispatch/dispatch_select_if.cuh
+++ b/cub/device/dispatch/dispatch_select_if.cuh
@@ -436,12 +436,12 @@ struct sm90_tuning<Input, true, true, 4, true, 4>
 template <class Input>
 struct sm90_tuning<Input, true, true, 4, true, 8>
 {
-    static constexpr int threads = 128;
-    static constexpr int items = 9;
+    static constexpr int threads = 224;
+    static constexpr int items = 6;
 
     static constexpr BlockLoadAlgorithm load_algorithm = BLOCK_LOAD_DIRECT;
 
-    using delay_constructor = detail::fixed_delay_constructor_t<468, 1175>;
+    using delay_constructor = detail::fixed_delay_constructor_t<532, 1180>;
 };
 
 #if CUB_IS_INT128_ENABLED