Nice read! Cool to see more GPU programming models that expose chiptet topology instead of treating it as a flat execution. Reminds me a little of Cerebras' CSL although far from as extreme as that.