My understanding from the article is that the code gen will still be deterministic--independent operations will be performed in parallel instead of in sequence, but data dependency will still be respected.