Eidola Simulator Models Multi-GPU Traffic for AI Workloads
New gem5 extension captures irregular GPU communication patterns with cycle-level precision.
As distributed AI training scales to hundreds of GPUs, techniques like kernel fusion and overlapping communication with computation introduce irregular, transient traffic patterns that existing simulators struggle to capture. Eidola, developed by researchers at the University of Wisconsin-Madison and AMD, extends the popular gem5 simulation framework to address this gap. The key innovation is its use of an "eidolon"—a succinct GPU model that retains only the minimal characteristics needed for accurate traffic modeling. By consuming annotated timing profiles from real GPU workloads, Eidola emulates peer-to-peer writes between GPUs with cycle-level accuracy, enabling researchers to study fine-grained synchronization behaviors across large configurations.
Eidola supports configurable per-GPU traffic patterns and allows isolated performance analysis under different communication scenarios. The team demonstrated its effectiveness by reproducing variability observed in fused kernel execution and by implementing a SyncMon-inspired synchronization mechanism that reduced polling-related memory traffic. The simulator provides a flexible platform for exploring interconnect bandwidth and latency trade-offs in distributed GPU systems. With 13 pages and 11 figures, the paper (arXiv:2606.12638) offers a practical tool for researchers and engineers designing next-generation AI hardware and networks.
- Eidola extends the gem5 framework with a minimalist GPU model (eidolon) for cycle-level traffic simulation.
- Uses annotated timing profiles from real applications to emulate peer-to-peer GPU writes accurately.
- Demonstrated reduction in polling-related memory traffic via a SyncMon-inspired synchronization mechanism.
Why It Matters
Provides a scalable simulation platform for optimizing interconnect design in large-scale multi-GPU AI systems.