What's New

Amazon EC2 G6 instances are now generally available on SageMaker Studio notebooks in the Middle East (Dubai) and Asia Pacific (Malaysia) regions. These instances are powered by NVIDIA L4 Tensor Core GPUs and AMD EPYC processors, delivering twice the deep learning inference performance of the previous-generation G4dn instances. This expansion enables customers in these regions to run interactive generative AI, NLP, computer vision, and recommender engine workloads directly within SageMaker Studio's JupyterLab and CodeEditor environments.

How It Works

• G6 instances are backed by up to 8 NVIDIA L4 Tensor Core GPUs, each with 24 GB of GPU memory, providing up to 192 GB of total GPU memory per instance for large model workloads.
• The instances pair NVIDIA L4 GPUs with third-generation AMD EPYC processors, balancing high-throughput GPU compute with strong CPU performance for data preprocessing and orchestration tasks.
• Within SageMaker Studio, users select a G6 instance as the kernel/compute resource when launching JupyterLab or CodeEditor applications, enabling interactive notebook-based experimentation.
• The G6 instances support interactive model training workflows, allowing data scientists to iterate on model fine-tuning directly in a notebook environment without provisioning separate training jobs.
• Users can also interactively test model deployment scenarios, simulating inference workloads before committing to a production endpoint, reducing the iteration cycle for MLOps pipelines.

Why It's Important

• Customers in the Middle East (Dubai) and Asia Pacific (Malaysia) regions can now access high-performance GPU compute locally, reducing latency and addressing data residency requirements that previously forced them to use distant regions.
• The 2x inference performance improvement over G4dn instances means faster experimentation cycles for generative AI fine-tuning and inference testing, directly accelerating time-to-production for ML teams.
• With 24 GB of GPU memory per L4 GPU, practitioners can load and experiment with larger foundation models interactively in notebooks without hitting memory constraints common on older GPU generations.
• Regional availability lowers the barrier for enterprises in emerging AI markets in the Middle East and Southeast Asia to adopt cloud-native ML development workflows on SageMaker Studio.
• Supporting use cases like NLP, language translation, computer vision, and recommender engines in a single instance family simplifies infrastructure decisions for teams running diverse ML workloads.

How It's Different

• G6 instances deliver 2x better deep learning inference performance compared to EC2 G4dn instances, which were powered by NVIDIA T4 GPUs with 16 GB of memory per GPU — a meaningful generational leap.
• The NVIDIA L4 Tensor Core GPU in G6 instances is purpose-built for inference efficiency and supports FP8 precision, enabling faster and more energy-efficient inference than the T4's FP16/INT8 capabilities.
• G6 instances offer 24 GB of GPU memory per card versus 16 GB on G4dn, allowing larger models to be loaded without multi-GPU tensor parallelism hacks during interactive experimentation.
• Pairing L4 GPUs with third-generation AMD EPYC processors (vs. Intel Cascade Lake on G4dn) provides improved CPU throughput for data pipelines running alongside GPU workloads in the same instance.
• Unlike dedicated training instances (e.g., P-series), G6 on SageMaker Studio is optimized for the interactive, iterative notebook workflow, bridging the gap between exploration and production deployment testing.

When to Prefer It

• Choose G6 instances when fine-tuning large language models or diffusion models interactively in SageMaker Studio notebooks, where the 24 GB per GPU memory headroom is critical for loading model weights.
• Prefer G6 over G4dn when your primary workload is deep learning inference testing, as the 2x performance improvement significantly reduces the time to evaluate model latency and throughput characteristics.
• Use G6 instances when operating in the Middle East (Dubai) or Asia Pacific (Malaysia) regions and data sovereignty or compliance requirements mandate that compute remain within those geographies.
• Select G6 for computer vision or NLP workloads that require rapid iteration between training runs and inference validation, leveraging the interactive SageMaker Studio environment for end-to-end experimentation.
• Prefer G6 when building and testing recommender engine models that involve both heavy embedding lookups (CPU-bound) and neural network scoring (GPU-bound), benefiting from the balanced AMD EPYC + L4 architecture.
• Consider G6 instances for generative AI prototyping workflows where teams need to quickly validate fine-tuned model outputs before promoting to a SageMaker Inference endpoint.

Availability

• Status: Generally Available (GA) as of May 11, 2026.
• New Regions: Middle East (Dubai) and Asia Pacific (Malaysia) — expanding the existing G6 availability footprint on SageMaker Studio notebooks.
• Access Method: Available through SageMaker Studio via JupyterLab and CodeEditor applications; refer to the SageMaker Studio developer guides for setup instructions.
• Pricing: Instance pricing follows standard EC2 G6 on-demand rates for SageMaker Studio; consult the SageMaker pricing page for region-specific costs.
• GPU Configuration: Supports configurations up to 8 NVIDIA L4 Tensor Core GPUs with 24 GB memory per GPU (up to 192 GB total GPU memory).
• Limitations: No specific limitations were disclosed in the announcement; availability may vary by specific G6 instance size (e.g., ml.g6.xlarge through ml.g6.48xlarge) within each region.