What's New

Amazon EC2 G6e instances are now generally available on SageMaker Studio notebooks in six additional regions: Middle East (Dubai), Asia Pacific (Tokyo, Seoul), and Europe (Frankfurt, Stockholm, Spain). These instances feature up to 8 NVIDIA L40S Tensor Core GPUs with 48 GB of GPU memory each, backed by third-generation AMD EPYC processors. The expansion enables developers in these regions to interactively fine-tune generative AI models, deploy LLMs up to 13B parameters, and generate images, video, and audio using diffusion models directly within SageMaker Studio's JupyterLab and CodeEditor environments.

How It Works

• G6e instances are powered by up to 8 NVIDIA L40S Tensor Core GPUs, each with 48 GB of dedicated GPU memory, providing up to 384 GB of total GPU memory per instance for memory-intensive workloads.
• Third-generation AMD EPYC processors handle CPU-side computation, complementing the GPU workloads for data preprocessing, orchestration, and mixed compute tasks.
• Users select G6e instance types when launching JupyterLab or CodeEditor applications within SageMaker Studio, enabling interactive notebook-based workflows without separate cluster provisioning.
• The instances support interactive model training workflows, allowing data scientists to iteratively fine-tune generative AI models and observe results in real time within the notebook environment.
• For inference testing, G6e instances can host LLMs with up to 13B parameters and run diffusion model inference for multimodal content generation (images, video, audio) directly from the notebook.

Why It's Important

• Developers in major financial, enterprise, and research hubs — Tokyo, Seoul, Frankfurt, Stockholm, Madrid, and Dubai — can now access high-performance GPU compute without routing workloads to distant regions, reducing latency and addressing data residency requirements.
• The 2.5x performance improvement over G5 instances means faster iteration cycles for generative AI fine-tuning, directly reducing the time-to-insight for ML practitioners.
• Having GPU-backed interactive notebooks lowers the barrier to entry for generative AI experimentation, allowing teams to prototype, fine-tune, and test LLMs and diffusion models in a single, managed environment.
• Regional availability supports compliance and sovereignty requirements for organizations in regulated industries (finance, healthcare, government) in Europe and the Middle East that cannot move data across geographic boundaries.
• The ability to interactively test model deployment on the same instance type used in production reduces the risk of performance surprises when moving from experimentation to deployment.

How It's Different

• G6e instances deliver up to 2.5x better performance than the previous-generation G5 instances (which use NVIDIA A10G GPUs), making them significantly more capable for large-scale generative AI workloads.
• The NVIDIA L40S GPU offers 48 GB of memory per GPU compared to 24 GB on the A10G in G5 instances, enabling larger models and larger batch sizes without memory bottlenecks.
• Unlike general-purpose GPU instances, G6e is specifically optimized for both training and inference of modern generative AI workloads, including transformer-based LLMs and diffusion models.
• The integration into SageMaker Studio notebooks differentiates this from raw EC2 access by providing a fully managed, IDE-like experience with built-in kernel management, experiment tracking, and AWS service integrations.
• The combination of AMD EPYC CPUs and NVIDIA L40S GPUs offers a heterogeneous compute profile that balances cost-efficiency with raw GPU throughput compared to purely NVIDIA-CPU-paired alternatives.

When to Prefer It

• Choose G6e on SageMaker Studio when fine-tuning open-source LLMs (e.g., Llama, Mistral) with up to 13B parameters interactively, where rapid iteration and real-time feedback are critical.
• Prefer G6e when working with diffusion models for image, video, or audio generation tasks that require large GPU memory buffers to hold model weights and intermediate activations.
• Use G6e when your organization operates under data residency or compliance requirements that mandate compute remain within specific regions such as EU (Frankfurt, Stockholm, Spain), Middle East (Dubai), or APAC (Tokyo, Seoul).
• Select G6e over G5 when workloads are GPU memory-bound and previously required model sharding or quantization workarounds due to the 24 GB per-GPU limit of G5 instances.
• Opt for G6e in SageMaker Studio when you need to interactively validate model deployment behavior before committing to a full SageMaker Endpoint deployment, reducing wasted inference infrastructure costs.
• Consider G6e for multimodal generative AI prototyping pipelines that combine text, image, and audio generation in a single notebook session requiring sustained high GPU throughput.

Availability

• Status: Generally Available (GA) as of May 11, 2026.
• New Regions: Middle East (Dubai), Asia Pacific (Tokyo), Asia Pacific (Seoul), Europe (Frankfurt), Europe (Stockholm), Europe (Spain).
• Supported Environments: Available for JupyterLab and CodeEditor applications within Amazon SageMaker Studio notebooks.
• Instance Specs: Up to 8 NVIDIA L40S Tensor Core GPUs, 48 GB GPU memory per GPU (up to 384 GB total), third-generation AMD EPYC processors.
• Pricing: Region-specific pricing is available on the AWS SageMaker pricing page; costs vary by instance size and region.
• Limitations: Support is scoped to SageMaker Studio notebook applications; availability for SageMaker Training Jobs or Inference Endpoints in these regions may differ and should be verified separately.