The new Amazon EC2 P6-B300 instances are designed to enhance the training and serving of large-scale AI models, such as those using Mixture of Experts (MoE) and multimodal processing techniques. These instances deliver a robust infrastructure by providing 6.4Tbps Elastic Fabric Adapter (EFA) networking, 2.1TB of GPU memory, and 192 VCPUs, allowing for efficient communication and reduced overhead. With 8x B300 GPUs and impressive data rates, they cater to organizations managing trillions of parameters requiring distributed training over thousands of GPUs.
Customers can benefit from the high-performance capabilities of P6-B300 for AI workloads, utilizing options like Amazon FSx for Lustre and Amazon S3 Express One Zone to enhance storage performance. The instances are specifically optimized for environments that demand quick accessibility for large models. Now available in the US West (Oregon) AWS Region, users can acquire these instances through various pricing models, highlighting the efficiency and scale of the AWS infrastructure for machine learning tasks.
👉 Pročitaj original: AWS Blog
