This article announces the general availability of Amazon EC2 P5 instances powered by NVIDIA H100 Tensor Core GPUs for accelerating generative AI and HPC applications.

• P5 instances deliver up to 6x faster training time and 40% lower training costs versus previous GPU instances
• Each instance includes 8 NVIDIA H100 GPUs with 640GB GPU memory and 3200 Gbps network bandwidth
• Supports up to 20,000+ H100 GPUs in EC2 UltraClusters with 20 exaflops aggregate compute capability
• Ideal for training large language models, computer vision, and demanding generative AI applications
• Includes second-generation Elastic Fabric Adapter for 8x better networking than P4d instances
• Available in US East (N. Virginia) and US West (Oregon) regions
• Integrates with SageMaker, AWS Batch, and ParallelCluster for managed deployment

EC2 P5 instances provide significant performance improvements for AI/ML and HPC workloads, enabling faster model training and reduced operational costs.