SageMaker Notebook Instances now support G6e instance types
G6e's 48 GB-per-GPU L40S hardware is now available in SageMaker notebooks, enabling interactive fine-tuning of 13B-parameter LLMs at 2.5x G5 speed.
View original announcement →Visual Summary
graph TD
A{{G6e on SageMaker Notebooks}}:::announced
B((ML Practitioner)):::external
C(NVIDIA L40S GPUs):::compute
D(SageMaker Notebook Instance):::compute
E([GenAI Fine-tuning]):::feature
F([LLM Deployment up to 13B]):::feature
G([Diffusion Models]):::feature
H(AWS Nitro System):::compute
I(Local NVMe SSD Storage):::storage
B ==>|"launches"| A
A ==>|"provisions"| D
D -->|"powered by"| C
D -->|"runs on"| H
D -->|"uses"| I
A -->|"enables"| E
A -->|"enables"| F
A -.->|"supports"| G
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
Amazon EC2 G6e instances are now generally available on SageMaker Notebook Instances, bringing NVIDIA L40S Tensor Core GPU-powered compute directly into the interactive notebook environment. These instances offer up to 8 GPUs with 48 GB of memory each, backed by third-generation AMD EPYC processors, delivering up to 2.5x better performance than the previous G5 generation. Customers can use them for interactive generative AI fine-tuning, LLM deployment up to 13B parameters, and diffusion model workloads for image, video, and audio generation.
How It Works
- SageMaker Notebook Instances provision a managed EC2 instance running a Jupyter Notebook server; selecting a G6e instance type now routes that provisioning to NVIDIA L40S-backed hardware.
- Each G6e instance features up to 8 NVIDIA L40S Tensor Core GPUs, each with 48 GB of GPU memory, totaling up to 384 GB of aggregate GPU memory per instance for memory-intensive model workloads.
- The underlying hardware uses fourth-generation tensor cores and third-generation AMD EPYC CPUs, supporting up to 192 vCPUs, 1.536 TB of system memory, 400 Gbps network bandwidth, and up to 7.6 TB of local NVMe SSD storage.
- G6e instances are built on the AWS Nitro System, which passes GPUs through in bare-metal mode, ensuring near-native GPU performance with no hypervisor overhead.
- Within the notebook environment, users can leverage standard NVIDIA libraries (TensorRT, CUDA, cuDNN) and SageMaker pre-installed deep learning frameworks (PyTorch, TensorFlow) to run training and inference experiments interactively.
- The same notebook instance can be used to test model deployment configurations before promoting to production SageMaker endpoints, enabling a tight iterate-and-validate loop.
Why It's Important
- Generative AI fine-tuning and LLM experimentation previously required spinning up separate training jobs or clusters; G6e notebook instances allow practitioners to do this interactively, dramatically shortening the iteration cycle.
- The 48 GB per-GPU memory footprint makes it feasible to load and experiment with models up to 13B parameters in a single notebook session without multi-node orchestration complexity.
- Diffusion model workloads (image, video, audio generation) are memory-bandwidth-bound; G6e's 2.9x faster GPU memory bandwidth over G6 instances directly translates to faster generation times during prototyping.
- Cost efficiency improves because G6e instances offer up to 20% cost savings over P4d instances while providing comparable or superior GPU memory, making high-end GPU experimentation more accessible.
- Spatial computing use cases (3D simulations, digital twins via NVIDIA Omniverse) gain a managed notebook entry point, broadening the addressable workload types within SageMaker.
How It's Different
- Compared to G5 instances, G6e delivers up to 2.5x better overall performance and 2x higher GPU memory per GPU (48 GB vs. 24 GB), enabling larger models and batch sizes in the same interactive environment.
- Compared to G6 instances, G6e provides 2x higher GPU memory and 2.9x faster GPU memory bandwidth, making it the superior choice for memory-bandwidth-sensitive inference and fine-tuning tasks.
- Compared to P4d instances, G6e offers up to 1.2x more total GPU memory and equivalent 400 Gbps networking at up to 20% lower cost, shifting the price-performance balance in favor of G6e for single-node workloads.
- Unlike P4d (which uses NVIDIA A100 GPUs optimized for large-scale distributed training), G6e with L40S GPUs combines AI compute with ray tracing and media acceleration, making it uniquely suited for both ML and spatial computing in one instance family.
- G6e is the first instance family in SageMaker Notebook Instances to support NVIDIA Omniverse-class spatial computing workloads, a capability not available in prior notebook instance GPU options.
When to Prefer It
- Use G6e notebook instances when interactively fine-tuning large language models (7B–13B parameters) where the full model and optimizer states must fit in GPU memory without model parallelism.
- Prefer G6e when prototyping or benchmarking diffusion models for image, video, or audio generation, where GPU memory bandwidth is the primary performance bottleneck.
- Choose G6e for iterative prompt engineering and inference testing on large models before committing to a production endpoint, leveraging the interactive notebook environment for rapid experimentation.
- Select G6e when your workload involves spatial computing tasks such as 3D simulation, digital twin construction, or NVIDIA Omniverse-based rendering alongside ML workflows in the same environment.
- Use G6e as a cost-effective alternative to P4d when running single-node fine-tuning or training of smaller models where 400 Gbps networking and high GPU memory are required but full multi-node scale is not.
- Prefer G6e over G5 whenever your model or dataset size has outgrown the 24 GB per-GPU memory limit of G5, or when you need meaningfully faster iteration times on existing G5 workloads.
Availability
- Status: Generally Available (GA) as of June 23, 2026.
- Supported Regions: US East (N. Virginia), US East (Ohio), US West (Oregon), Asia Pacific (Tokyo), Middle East (Dubai), Europe (Frankfurt), Europe (Sweden), Europe (Spain).
- Pricing Model: Standard SageMaker Notebook Instance pricing applies; billed per-second for the duration the instance is running, at the EC2 G6e on-demand rate plus SageMaker overhead (refer to the SageMaker Pricing page for current rates).
- Integration: Available via the SageMaker console, AWS CLI, and SageMaker SDK when creating or updating a notebook instance; also accessible through JupyterLab and Code Editor applications in SageMaker Studio.
- Limitation: Not yet available in all AWS regions; customers in regions outside the listed eight will need to use alternative GPU instance types or request access through AWS Support.
- Maintenance: SageMaker AI updates the underlying notebook instance software at least every 90 days; AWS recommends restarting notebook instances at least every 30 days to consume security patches.
Related Resources
- https://aws.amazon.com/ec2/instance-types/g6e/
- https://docs.aws.amazon.com/sagemaker/latest/dg/studio-updated-jl.html
- https://docs.aws.amazon.com/sagemaker/latest/dg/code-editor.html
- https://docs.aws.amazon.com/sagemaker/latest/dg/studio-updated.html
- https://docs.aws.amazon.com/sagemaker/latest/dg/nbi.html
