Amazon Launches EC2 P6-B200 Instances to Boost AI Capabilities with NVIDIA Blackwell GPUs

Jacob Heinz

Jacob Heinz

— 2 min read
NVIDIA Blackwell GPU inside a futuristic data center

Amazon Web Services (AWS) has unveiled its latest offering, the Amazon EC2 P6-B200 instances, which are powered by NVIDIA's cutting-edge Blackwell GPUs. This new instance type is designed to enhance performance and scalability for artificial intelligence (AI), machine learning (ML), and high-performance computing (HPC) applications, catering to the growing demand for advanced computational resources.

Key Takeaways

  • Enhanced Performance: P6-B200 instances deliver up to 2x the performance for AI training and inference compared to previous models.
  • Advanced Specifications: Each instance features eight NVIDIA B200 GPUs, 1440 GB of GPU memory, and 5th Generation Intel Xeon processors.
  • Flexible Capacity Reservations: Users can reserve EC2 Capacity Blocks for ML workloads for periods ranging from 1 to 182 days.
  • Integration with AWS Services: Seamless compatibility with AWS managed services like Amazon EKS and Amazon S3.

Overview of EC2 P6-B200 Instances

The new EC2 P6-B200 instances are specifically tailored for large-scale distributed AI training and inference, making them ideal for tasks such as:

  • Foundation model training with reinforcement learning and distillation.
  • Multimodal training and inference.
  • High-performance computing applications, including climate modeling and drug discovery.

These instances are equipped with:

  • 8 NVIDIA B200 GPUs
  • 1440 GB of high bandwidth GPU memory
  • 5th Generation Intel Xeon Scalable processors (Emerald Rapids)
  • 2 TiB of system memory
  • 30 TB of local NVMe storage

Performance Improvements

The P6-B200 instances boast significant enhancements over their predecessors, the EC2 P5en instances:

  • 125% improvement in GPU TFLOPs
  • 27% increase in GPU memory size
  • 60% increase in GPU memory bandwidth

This performance leap allows users to accelerate their time-to-market for AI applications and reduces costs associated with inference.

Capacity Reservations and Usage

Users can access the P6-B200 instances in the US West (Oregon) AWS Region through EC2 Capacity Blocks. The process for reserving capacity is straightforward:

  1. Navigate to the Amazon EC2 console.
  2. Select "Purchase Capacity Blocks for ML".
  3. Specify the total capacity and duration (1-182 days).
  4. Choose an earliest start date up to 8 weeks in advance.

The total price for the EC2 Capacity Block is charged upfront, ensuring predictable costs for users.

Integration with AWS Ecosystem

The P6-B200 instances can be easily integrated with various AWS services, enhancing their utility:

  • Amazon Elastic Kubernetes Services (EKS) for container orchestration.
  • Amazon Simple Storage Service (S3) for scalable storage solutions.
  • Amazon FSx for Lustre for high-performance file systems.

Support for Amazon SageMaker HyperPod is also on the horizon, further expanding the capabilities of these instances for machine learning practitioners.

Conclusion

The launch of Amazon EC2 P6-B200 instances marks a significant advancement in cloud computing capabilities, particularly for AI and ML applications. With their powerful specifications and enhanced performance, these instances are set to accelerate innovation and drive the adoption of generative AI technologies across various industries. Users can start utilizing these instances today through the Amazon EC2 console, paving the way for faster and more efficient AI development.

Sources

Read more