AI @Scale
summary
Microsoft describes a unified Azure AI @Scale stack that actually ships: ACPT containers bundle PyTorch with DeepSpeed and Onnx/Onyx runtime so teams run the same validated binaries on Nvidia A100 and AMD GPUs. Nebula implements three-tier checkpointing (device→host→peer→remote) and checkpoints a 530B-parameter model across 22,000 A100s in about 20 seconds. MSQL generates optimal collective-communication algorithms and ONNX execution providers plus a model optimization service produce hardware-targeted compiled packages for cloud+edge deployment.
AI-written summary · about
chapters (6)
- 0:00Microsoft AI scale history
- 4:04Supercomputer infrastructure
- 4:51Software training stack
- 14:54Hybrid inference loop
- 19:05Hardware abstraction layers
- 20:08Future and composition
related talks
50mAccelerate development and deployment of Gen AI with NVIDIA on Azure
24mHow to build next-gen AI services with NVIDIA AI on Azure Cloud
28mHarnessing generative AI with NVIDIA AI and Microsoft Azure
14mAnyscale on Azure: Build, Train, and Serve AI in Your Own Tenant
28mScaling agentic AI: From infrastructure to enterprise outcomes
25mOptimize productivity by meeting next-level demands of the AI era
40mBuilding secure and efficient AI solutions in Azure with Partners
06mPower today's most sophisticated models with AMD Instinct MI300X