Ankur Srivastava

Performance optimization of distributed training jobs is complex due to the multiple levels of parallelism and the need to drill down in the accelerator, CPU, and network metrics at nanosecond granularity throughout the stack, from the communication libraries driving collective network operations to the ML framework.

SaaS is the preferred consumption model for many enterprises because it optimizes costs while improving speed and deployment agility. However, some organizations may struggle with SaaS configuration and vendor compliance assessments.

Attend this chalk talk with Meta PyTorch and AWS to learn about distributed training, accelerated inference, bin-packing models, and scaling with PyTorch on AWS. In addition, hear about relevant large-companies that use AWS infrastructure and services for distributed training and to serve thousands of models with low latency and cost.

Discover the power of combining traditional HPC simulations with AI/ML services to activate new workflows and scientific breakthroughs. In this chalk talk, learn about combined AWS HPC and ML solutions that have evolved to state-of-the-art supercomputing models, powered by AWS services like Elastic Fabric Adapter, AWS ParallelCluster, Amazon EC2 instances, Amazon FSx for Lustre, and AWS Batch.