diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 4021113f5e..2fc42190f3 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -9078,7 +9078,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     uint32_t split_k = 1;
 
     // Intel Alchemist prefers more workgroups
-    const uint32_t shader_core_count_multiplier = (ctx->device->vendor_id == VK_VENDOR_ID_INTEL && ctx->device->architecture != INTEL_XE2) ? 2 : 1;
+    const uint32_t shader_core_count_multiplier = (ctx->device->vendor_id == VK_VENDOR_ID_INTEL && ctx->device->architecture != INTEL_XE2_ONWARD) ? 2 : 1;
 
     // Use a placeholder core count if one isn't available. split_k is a big help for perf.
     const uint32_t shader_core_count = ctx->device->shader_core_count ? ctx->device->shader_core_count * shader_core_count_multiplier : 16;