What's New

Amazon EC2 G6 instances are now generally available on SageMaker notebook instances across nine additional regions in Asia Pacific and Europe. These instances are powered by NVIDIA L4 Tensor Core GPUs and offer twice the deep learning inference performance compared to the previous-generation G4dn instances. The expansion enables data scientists and ML engineers in these regions to interactively develop, fine-tune, and test generative AI and other ML workloads directly within SageMaker's notebook environment.

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 memory-intensive model workloads.
• The underlying CPU is a third-generation AMD EPYC processor, offering a high-performance host compute layer to complement the GPU resources.
• Users select a G6 instance type when creating or updating a SageMaker notebook instance, and the environment is accessible via JupyterLab or CodeEditor applications within SageMaker Studio.
• The instances support interactive workflows, meaning users can iteratively write code, run training loops, test inference endpoints, and inspect outputs in real time within the notebook interface.
• G6 instances leverage the NVIDIA L4's INT8 and FP16 Tensor Core capabilities, which are specifically optimized for inference and fine-tuning tasks common in modern generative AI pipelines.

Why It's Important

• Expanding to nine new regions reduces latency and data residency concerns for customers in Asia Pacific and Europe who previously had to route workloads to other regions to access G6 hardware.
• The 2x inference performance improvement over G4dn translates directly into faster iteration cycles during interactive development, reducing the time between code changes and validated results.
• Generative AI fine-tuning and inference are among the most GPU-memory-hungry workloads; the 24 GB per GPU on L4 makes it feasible to load larger models (e.g., 7B–13B parameter LLMs) without multi-instance workarounds.
• Availability in SageMaker notebook instances lowers the barrier to entry—teams can prototype on the same GPU class they intend to use in production without managing raw EC2 infrastructure.
• Broader regional availability supports enterprise compliance requirements, particularly for European customers subject to GDPR data locality obligations.

How It's Different

• G6 instances use NVIDIA L4 GPUs rather than the NVIDIA T4 GPUs found in G4dn instances, delivering 2x better deep learning inference throughput at comparable or lower cost per inference operation.
• The L4 GPU introduces fourth-generation Tensor Cores with improved sparsity support, enabling more efficient execution of pruned and quantized models compared to T4-based G4dn instances.
• Each L4 GPU carries 24 GB of GDDR6 memory versus 16 GB on the T4, allowing larger models or larger batch sizes to fit entirely on a single GPU without offloading.
• G6 pairs the NVIDIA L4 with AMD EPYC CPUs rather than the Intel Cascade Lake CPUs in G4dn, offering higher core counts and memory bandwidth on the host side.
• Unlike G4dn, which was primarily positioned for inference at launch, G6 is explicitly supported for both interactive training (including generative AI fine-tuning) and inference within SageMaker notebooks.

When to Prefer It

• Choose G6 when fine-tuning large language models (e.g., using LoRA or QLoRA techniques) interactively in a notebook, where the 24 GB per GPU enables loading 7B–13B parameter models on a single GPU.
• Prefer G6 over G4dn when running real-time inference tests during development and iteration speed is critical, given the 2x inference performance advantage.
• Use G6 for computer vision workloads involving large batch image processing or video frame inference, where the higher GPU memory and Tensor Core throughput reduce per-batch latency.
• Select G6 for NLP and language translation prototyping when models exceed the 16 GB memory ceiling of G4dn's T4 GPU, avoiding the complexity of model sharding across multiple GPUs.
• Opt for G6 in European or Asia Pacific regions when data residency regulations prohibit sending training data or model weights to other AWS regions where G6 was previously available.
• Use G6 for recommender engine development when embedding tables and interaction matrices are large enough to benefit from the expanded GPU memory and faster matrix operations.

Availability

• Status: Generally available (GA) as of May 12, 2026.
• New Regions: Asia Pacific (Tokyo, Mumbai, Sydney) and Europe (London, Paris, Frankfurt, Stockholm, Zurich).
• Service: Available on SageMaker notebook instances; accessible via JupyterLab and CodeEditor applications in SageMaker Studio.
• Pricing: Billed at standard Amazon EC2 G6 on-demand instance rates plus SageMaker notebook instance overhead; specific pricing varies by region and instance size—consult the SageMaker pricing page for regional rates.
• Limitations: No specific limitations were disclosed in the announcement; standard SageMaker notebook instance quotas and service limits apply per account and region.
• Documentation: Setup instructions are available in the SageMaker developer guides for JupyterLab and CodeEditor applications.