---

Summary

The NVIDIA Grace-Hopper Superchip tightly couples CPU and GPU memory systems, but the consistency and coherence behavior exposed to software is not fully captured by traditional CPU-only or GPU-only memory-model assumptions. This work characterizes the ordering and coherence properties visible across the Grace CPU and Hopper GPU.

The paper uses litmus tests and microbenchmarks to identify the practical semantics that system software and runtimes can rely on when sharing data across the CPU-GPU boundary.

Bibtex

@inproceedings{bagchi:grace-hopper,
  author = {Soham Bagchi and Sanya Srivastava and Reese Levine and Tyler Sorensen and Ryan Stutsman and Vijay Nagarajan},
  title = {{Consistency and Coherence of the NVIDIA Grace-Hopper Superchip}},
  booktitle = {Proceedings of the 2026 ACM SIGPLAN International Symposium on Memory Management},
  series = {ISMM '26},
  year = {2026},
  pages = {66--79},
  publisher = {ACM},
  url = {https://doi.org/10.1145/3814942.3816134},
  doi = {10.1145/3814942.3816134},
}

---