Workshop: PMBS26: The 17th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems
Authors: Rubén Coll Sánchez (Sorbonne University, Huawei Technologies France SASU); Thibaut Tachon and Pierre Leca (Huawei Technologies France SASU); Teng Su (Huawei Technologies Co. Ltd); and Chong Li (Huawei Technologies France SASU)
Abstract: Developing large-scale AI systems requires extensive on-machine tuning: engineers evaluate hundreds of parallel configurations by trial and error, at considerable cost in accelerator rental and compute time. Existing performance models are either tied to specific frameworks or too coarse for modern distributed workloads. We present AMAI, a performance analysis framework featuring (i) a declarative domain-specific language for describing distributed deep-learning computations, and (ii) compositional cost semantics that map programs into symbolic models of time and per-device memory. Models are parameterized by hardware constants and structural variables that define the parallelism configuration. As a result, descriptions can be analyzed at any scale and reused across machines. Predictions stay within 3 percent of measured step time and peak memory across batch-size and data-parallel scaling when analyzing the training of Qwen3.5 0.8B on a multi-node NPU cluster.
Back to PMBS26: The 17th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems Archive Listing
Back to Full Workshop Archive Listing