Search papers, labs, and topics across Lattice.
This paper introduces Eidola, an extension to the gem5 simulation framework designed to model inter-GPU communication traffic in distributed AI workloads. By leveraging annotated timing profiles from real applications, Eidola achieves cycle-level precision in emulating peer-to-peer GPU writes, allowing for detailed analysis of synchronization behavior across large multi-GPU configurations. The results demonstrate Eidola's capability to reproduce variability in fused kernel execution and reduce polling-related memory traffic, highlighting its potential for architectural exploration in modern GPU systems.
Eidola reveals that precise modeling of inter-GPU communication can drastically reduce memory traffic and enhance performance in distributed AI systems.
As distributed AI workloads grow in scale, multi-GPU systems have become essential for training large models. Although techniques like kernel fusion and overlapping communication with computation help reduce delays, they also introduce irregular and transient traffic patterns that are difficult to model using existing tools. These techniques rely heavily on fine-grained synchronization and peer-to-peer communication, which place significant pressure on interconnect bandwidth and latency. In this work, we introduce Eidola, a scalable extension to the gem5 simulation framework that enables detailed modeling of inter-GPU communication traffic. The extension is scalable as our GPU model serves as a succinct eidolon, emulating the minimal characteristics needed for traffic modeling. Eidola uses annotated timing profiles from real applications to emulate peer-to-peer GPU writes with cycle-level precision. This allows researchers to simulate and analyze synchronization behavior across large multi-GPU configurations. The simulator supports configurable per-GPU traffic patterns and enables isolated performance analysis under different communication scenarios. We demonstrate Eidola's effectiveness by reproducing variability in fused kernel execution and by implementing a SyncMon-inspired synchronization mechanism, confirming reductions in polling-related memory traffic. Our results show that Eidola provides a flexible and scalable platform for studying inter-GPU communication and supports architectural exploration in modern distributed GPU systems.
