SimAI: Unifying Architecture Design and Performance Tuning for Large-Scale Large...
summary
SimAI claims a unified, high‑precision LLM training simulator that matches real runs 98.1% alignment by hijacking training frameworks (Megatron, DeepSpeed), profiling kernels on real GPUs, and replacing collectives with an enhanced NS3 model via SimCCL collective‑to‑point‑to‑point translation. It implements lock‑free global context‑sharing and multi‑thread acceleration, runs single‑host workload generation plus kernel DB lookup, validates on 128‑GPU testbeds (H100 vs A100), keeps intrahost error <4%, and reports 36× end‑to‑end accuracy improvement over AstraSim; open source.
AI-written summary · about
chapters (8)
- 0:00SimAI introduction
- 1:45Precise training simulator challenges
- 2:51SimAI architecture overview
- 3:55Workload and kernel simulation
- 4:59Collective communication with SimCCL
- 6:23Experimental evaluation results
- 8:13Production use cases
- 9:41Future work and open-source
related talks
32mPanel Discussion: The AI Kernel...
25mConvergence of AI, Simulations and HPC
22mCentimani: Enabling Fast AI Accelerator Selection for DNN Training with a Novel...
24mKeynote: Next-generation frameworks for Large-scale AI
22mGenAI Training In Production: Software, Hardware & Network Considerations
23mAutoSys: The Design and Operation of Learning-Augmented Systems
40mOptimize Your AI Cloud Infrastructure: A Hardware Perspective
32mDeploy AI with Simulink