Consistency and Coherence of the NVIDIA Grace-Hopper Superchip
summary
Empirical study validates the Grace‑Hopper enforces the Compound Memory Consistency Model using 1,960 litmus variants spanning ARM and PTX, showing independent compilation composes safely when scopes and synchronizations match. New value propagation tests reverse‑engineer coherence: GPU uses write‑through L1 with self‑invalidation rather than writer‑initiated invalidations, CPU‑GPU coherence uses AMBA CHI‑like directory invalidations. Surprising extra strength appears for read‑modify‑write atomics (stronger than PTX/CMCM spec); experiments and numbers are reproducible via the cited paper.
AI-written summary · about
chapters (8)
- 0:00Grace-Hopper consistency research
- 1:32Grace-Hopper architecture overview
- 3:21ARM-PTX memory model validation
- 9:26CPU and GPU memory models baseline
- 12:40Heterogeneous message passing tests
- 17:37Grace-Hopper coherence probing
- 23:56RMW behavior and performance
- 25:44Q&A
related talks
45mHackathon Presentations
01h 21mPanel+Closing
01h 15mThe Evolution of Memory Technologies
10mA Programming Model for Efficient Inter-Kernel Control-Flow on Memory-Mapped Near-Data(…)
01h 20mExtending Moore’s Law with Low-Cost Memory Reliability
10mOmniXtend: Open Coherent Memory Fabric for RISC-V
14mStrategies for CXL’s success in a world of proprietary Coherent...
01h 04mThe challenge of managing parallelism and data-movement for tensor computations on GPUs