Xsim simulator cuts LLM training simulation errors to under 5%

Xsim, a new heterogeneity-aware simulator from researchers including Sumit Kumar and eight co-authors, tackles the growing challenge of training large language models on non-uniform hardware. Real-world AI infrastructure is increasingly heterogeneous due to multimodal/MoE architectures, limited homogeneous GPU availability on public clouds, and geographically distributed deployments. Existing simulators assume homogeneous compute and networks, leading to inaccurate performance predictions. Xsim addresses this with load balancing via non-uniform workload partitioning across heterogeneous device groups, heterogeneity-aware collective communication through customized ring construction and chunk partitioning, and reusable abstractions for pipeline-parallel algorithms and non-uniform tensor resharding. It also offers flexible input abstractions for custom device groups and device-to-parallelism mappings, plus pluggable integration with NS-3 and htsim to trade simulation fidelity for scalability.

In evaluation, Xsim accurately predicts training time for real-world heterogeneous deployments, achieving less than 5% error across most heterogeneous data-parallel and tensor-parallel configurations, and around 2% error with pipeline-parallel communication modeling. It exposes actionable metrics such as pipeline bubble time and straggler waiting time, enabling engineers to identify bottlenecks. For professionals managing distributed AI training, Xsim provides a practical tool to plan and optimize deployments across diverse GPU clusters—critical for cutting costs and reducing time-to-train in increasingly heterogeneous environments.