It's probably a sweet spot between adding more CU's versus cache. Given their memory bandwidth is lower, alleviating that with cache makes more sense than adding more CU's.