Fleet: Hierarchical Task-Based Abstraction for Megakernels on Multi-Die GPUs

mmatt_d about 3 hours ago 1 commentsRead Article on arxiv.org

DE version is available. Content is displayed in original English for accuracy.

Discussion (1 Comments)Read Original on HackerNews

joshuakelleyds•about 2 hours ago

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.