What's New

Amazon EC2 P6-B200 instances are now generally available in SageMaker Studio notebooks in the US East (N. Virginia) region, expanding access to NVIDIA Blackwell GPU-powered compute for interactive ML development. These instances feature 8 NVIDIA B200 GPUs with 1,440 GB of high-bandwidth GPU memory paired with 5th Generation Intel Xeon (Emerald Rapids) processors. Customers can use them directly within JupyterLab or CodeEditor environments to develop, fine-tune, and experiment with large foundation models.

How It Works

• P6-B200 instances are provisioned as notebook kernels within SageMaker Studio, accessible through the JupyterLab or CodeEditor application interfaces without requiring separate EC2 cluster management.
• Each instance is equipped with 8 NVIDIA Blackwell B200 GPUs, providing 1,440 GB of aggregate high-bandwidth memory (HBM3e), enabling in-memory hosting of very large model parameter sets during interactive sessions.
• The underlying CPU is a 5th Generation Intel Xeon Scalable processor (Emerald Rapids), which handles data preprocessing, orchestration, and host-side compute tasks alongside the GPU workloads.
• Users interact with models and training loops in real time through notebook cells, leveraging frameworks such as PyTorch, JAX, or TensorFlow with CUDA/NCCL backends optimized for Blackwell architecture.
• SageMaker Studio manages instance lifecycle, IAM-based access control, and storage integration (e.g., Amazon EFS, S3), so researchers can focus on model experimentation rather than infrastructure provisioning.

Why It's Important

• The 1,440 GB of GPU memory allows practitioners to load and fine-tune frontier-scale models—including 70B+ parameter LLMs and mixture-of-experts architectures—entirely within a single interactive notebook session, eliminating the need to shard across multiple nodes for experimentation.
• Up to 2x training throughput improvement over P5en instances means iteration cycles are significantly shorter, accelerating the research-to-production pipeline for generative AI applications.
• Native integration with SageMaker Studio lowers the operational barrier: teams without deep MLOps expertise can access state-of-the-art GPU hardware through a managed, familiar notebook interface.
• Support for multi-modal reasoning models (text, image, video) broadens the range of enterprise use cases—such as copilots and content generation—that can be prototyped and validated interactively before committing to full training runs.

How It's Different

• P6-B200 instances deliver up to 2x better AI training performance compared to P5en (NVIDIA H100) instances, reflecting the architectural leap from Hopper to Blackwell GPU microarchitecture.
• The 1,440 GB aggregate GPU memory pool is substantially larger than the 640 GB available on an 8-GPU P5en instance, enabling larger batch sizes and eliminating gradient checkpointing trade-offs for many model sizes.
• NVIDIA Blackwell introduces 5th-generation NVLink and NVSwitch interconnects, providing higher GPU-to-GPU bandwidth than the prior generation, which benefits multi-GPU tensor-parallel workloads common in LLM fine-tuning.
• Unlike raw EC2 access, the SageMaker Studio integration provides built-in experiment tracking, managed kernels, and direct connectivity to AWS data services, reducing setup overhead compared to self-managed Blackwell instances.
• The Emerald Rapids CPU pairing offers improved PCIe 5.0 bandwidth and higher core counts versus the Sapphire Rapids CPUs in P5en, reducing CPU-side bottlenecks in data-intensive preprocessing pipelines.

When to Prefer It

• Choose P6-B200 Studio notebooks when interactively fine-tuning large foundation models (e.g., 70B–400B parameter LLMs or MoE models) where the full model must reside in GPU memory during experimentation.
• Prefer these instances for rapid prototyping of multi-modal generative AI applications—such as video generation or vision-language models—that require high memory capacity and compute throughput within a single session.
• Use P6-B200 when iteration speed is critical and the 2x performance gain over P5en meaningfully reduces the feedback loop between hypothesis and result during model development.
• Ideal for enterprise teams building copilots or domain-specific generative AI tools who need a managed, secure notebook environment without provisioning and configuring raw GPU clusters.
• Select this option when working with mixture-of-experts architectures that have large sparse parameter counts, as the expanded GPU memory accommodates expert routing and activation without offloading to CPU RAM.

Availability

• Status: Generally Available (GA) as of May 11, 2026.
• Supported Region: AWS US East (N. Virginia) — this announcement represents a region expansion, implying prior availability in at least one other region.
• Access Method: Available directly within SageMaker Studio via JupyterLab and CodeEditor application environments.
• Pricing: On-demand pricing applies; specific rates are listed on the SageMaker Studio pricing page (not disclosed in the announcement).
• Setup: Configuration instructions are available in the SageMaker Studio developer guides for JupyterLab and CodeEditor applications.
• Limitations: No additional regions beyond US East (N. Virginia) are confirmed in this announcement; availability in other regions has not been specified.