What's New

Amazon Web Services has expanded the availability of Amazon EC2 P4de instances on SageMaker Studio notebooks to three new regions: Asia Pacific (Tokyo), Asia Pacific (Singapore), and Europe (Frankfurt). These instances are equipped with 8 NVIDIA A100 GPUs, each featuring 80GB of HBM2e GPU memory, delivering a total of 640GB of GPU memory per instance. This expansion brings high-performance ML training capabilities closer to customers in Asia and Europe, enabling faster iteration and reduced latency for data-intensive workloads.

How It Works

• P4de instances are powered by 8 NVIDIA A100 Tensor Core GPUs, each with 80GB of HBM2e high-bandwidth memory, totaling 640GB of GPU memory per instance.
• HBM2e memory provides significantly higher memory bandwidth compared to previous generations, enabling faster data movement between GPU cores and memory during training operations.
• The instances integrate directly with SageMaker Studio notebooks, accessible via JupyterLab and CodeEditor applications, allowing data scientists to launch GPU-backed notebook kernels without managing underlying infrastructure.
• Users can configure and access P4de instances through the SageMaker Studio developer guides for both JupyterLab and CodeEditor environments, following standard instance selection workflows.
• The instances leverage NVLink and high-speed interconnects between GPUs to support large-scale distributed training jobs that require tight GPU-to-GPU communication.

Why It's Important

• The regional expansion reduces data residency and latency concerns for customers in Japan, Singapore, and Germany who must keep training workloads within specific geographic boundaries for compliance or regulatory reasons.
• Up to 60% better ML training performance compared to P4d instances means teams can iterate on model development significantly faster, compressing experiment cycles and accelerating time to market.
• A 20% lower cost to train relative to P4d instances means organizations get more compute value per dollar, making large-scale training more economically viable.
• The 640GB total GPU memory pool enables training of very large models or processing of high-resolution datasets that would otherwise require complex model parallelism workarounds or be infeasible on lower-memory instances.
• Availability within SageMaker Studio notebooks lowers the barrier to accessing this hardware tier, allowing data scientists to use it interactively without needing to configure standalone EC2 clusters.

How It's Different

• P4de instances offer 2X the per-GPU memory (80GB vs. 40GB) compared to P4d instances, directly enabling larger batch sizes, bigger model checkpoints, and higher-resolution input data without out-of-memory errors.
• The 60% ML training performance improvement over P4d translates to measurably shorter wall-clock training times for equivalent workloads, not just marginal gains.
• Despite the significantly higher memory and performance, P4de instances cost 20% less to train on than P4d, inverting the typical trade-off between capability and cost.
• HBM2e memory technology provides higher memory bandwidth than the HBM2 used in P4d instances, reducing memory bottlenecks in bandwidth-bound training scenarios such as large transformer models.
• Integration with SageMaker Studio differentiates this from raw EC2 access by providing managed notebook environments, experiment tracking, and seamless access to other SageMaker features alongside the high-performance hardware.

When to Prefer It

• Choose P4de when training large language models (LLMs) or foundation models that require more than 40GB of GPU memory per device and would otherwise require aggressive model sharding on P4d instances.
• Prefer P4de for computer vision workloads involving high-resolution imagery (e.g., medical imaging, satellite imagery) where large input tensors quickly exhaust lower-memory GPUs.
• Use P4de when operating in Asia Pacific (Tokyo), Asia Pacific (Singapore), or Europe (Frankfurt) and data sovereignty or latency requirements prevent routing workloads to other regions where P4de was previously available.
• Select P4de over P4d when training cost efficiency is a priority, as the 20% lower training cost combined with faster completion times yields better overall economics for long-running jobs.
• Opt for P4de in SageMaker Studio notebooks when data scientists need interactive, exploratory access to high-end GPU hardware without the overhead of provisioning and managing dedicated training clusters.
• Consider P4de for multi-GPU distributed training experiments where the higher per-GPU memory reduces the need for gradient checkpointing or other memory-saving techniques that can slow training.

Availability

• Status: Generally Available (GA) as of May 11, 2026.
• New Regions: Asia Pacific (Tokyo), Asia Pacific (Singapore), and Europe (Frankfurt).
• Access Method: Available through SageMaker Studio notebooks via JupyterLab and CodeEditor applications.
• Pricing: Region-specific pricing is available on the AWS SageMaker pricing page; the instances offer approximately 20% lower cost to train compared to P4d instances.
• Hardware Spec: Each P4de instance includes 8 × NVIDIA A100 GPUs with 80GB HBM2e memory each (640GB total GPU memory).
• Prerequisite: Users should refer to the SageMaker Studio developer guides for setup instructions specific to JupyterLab and CodeEditor environments.