diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h
index 953b71fb8a..8ae3a4b148 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h
@@ -292,7 +292,7 @@ struct __parallel_merge_submitter_large<_IdType, _CustomName,
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
 
         // Empirical number of values to process per work-item
-        const std::uint8_t __chunk = __exec.queue().get_device().is_cpu() ? 128 : 8;
+        const std::uint8_t __chunk = __exec.queue().get_device().is_cpu() ? 128 : 4;
 
         const _IdType __steps = oneapi::dpl::__internal::__dpl_ceiling_div(__n, __chunk);
         const _IdType __base_diag_count = 1'024 * 32;