SageMaker Notebook Instances now support P5.4xl instance types
H100 GPU power comes to SageMaker notebooks—fine-tune LLMs and diffusion models interactively with up to 4x faster performance.
View original announcement →Visual Summary
graph TD
A{{SageMaker P5.4xl Notebook Instances}}:::announced
B((Data Scientists)):::external
C(NVIDIA H100 GPU):::compute
D(SageMaker Studio):::compute
E([JupyterLab IDE]):::feature
F([Code Editor IDE]):::feature
G([LLM Fine-tuning]):::feature
H(Amazon EC2 P5):::compute
I(Amazon EBS Storage):::storage
B ==>|"launches"| A
A -->|"powered by"| C
A -->|"runs within"| D
D -->|"provides"| E
D -->|"provides"| F
A -->|"enables"| G
A -->|"built on"| H
A -->|"persists data"| I
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 SageMaker Notebook Instances now support the EC2 P5.4xl instance type, powered by NVIDIA H100 Tensor Core GPUs, making it generally available for interactive ML development. This enables data scientists and ML practitioners to leverage H100 GPU performance directly within Jupyter notebooks for training and experimenting with large language models, diffusion models, and other compute-intensive workloads. The addition brings the same high-performance GPU infrastructure previously available in EC2 into the familiar SageMaker notebook environment.
How It Works
- NVIDIA H100 Tensor Core GPU: The P5.4xl instance is equipped with an H100 GPU, delivering state-of-the-art tensor compute performance for deep learning training and inference tasks.
- SageMaker Notebook Instance integration: Users can select
ml.p5.4xlargeas the instance type when creating or updating a SageMaker Notebook Instance, launching a Jupyter server on the underlying EC2 P5.4xl hardware. - Preconfigured environment: The notebook instance comes with the SageMaker Python SDK, Boto3, deep learning framework libraries (PyTorch, TensorFlow, etc.), and Conda pre-installed, reducing setup overhead.
- Persistent storage: Data stored under
/home/ec2-user/SageMakeris persisted across instance restarts, while data outside this directory is not guaranteed to persist during maintenance events. - IDE compatibility: The P5.4xl instance can be used with JupyterLab and Code Editor applications within SageMaker Studio, as well as classic SageMaker Notebook Instances.
- EFA and UltraCluster context: While the P5.4xl is a single-instance form factor, the broader P5 family supports up to 3,200 Gbps networking via second-generation Elastic Fabric Adapter (EFA) for distributed workloads.
Why It's Important
- Faster experimentation on cutting-edge hardware: Researchers and engineers can now prototype, fine-tune, and validate LLMs and diffusion models on H100 GPUs directly within an interactive notebook, dramatically shortening the iteration cycle.
- Up to 4x performance improvement: Compared to previous-generation GPU instances (e.g., P4), the H100-based P5.4xl can reduce training and inference time by up to 4x, translating directly to faster time-to-insight.
- Up to 40% cost reduction for ML training: The performance-per-dollar improvement means organizations can train the same models at significantly lower cost, making advanced generative AI more economically accessible.
- Generative AI use cases unlocked interactively: Complex workloads such as LLM fine-tuning, image/video generation, code generation, and speech recognition can now be explored interactively without needing to set up full training jobs.
- Reduced operational friction: By surfacing P5.4xl within the managed SageMaker Notebook Instances experience, AWS handles infrastructure provisioning, kernel management, and software maintenance, letting practitioners focus on model development.
How It's Different
- H100 vs. previous-generation GPUs: Unlike P3 (V100) or P4 (A100) instances, the P5.4xl uses the NVIDIA H100 Tensor Core GPU, which offers significantly higher FP8/FP16/BF16 throughput and a larger HBM3 memory bandwidth, critical for modern LLM workloads.
- Interactive notebook access vs. training jobs only: Previously, H100-class compute was primarily accessible via SageMaker Training Jobs or EC2 directly; this announcement brings it into the interactive, managed notebook experience for the first time.
- 2x CPU and memory, 4x local storage vs. P3/P4: The P5 family complements the GPU with 2x higher CPU performance, 2x higher system memory, and 4x higher local storage compared to previous-generation GPU instances, benefiting data preprocessing pipelines run alongside model training.
- Managed environment vs. raw EC2: Unlike launching a P5.4xl on EC2 directly, SageMaker Notebook Instances provide pre-installed ML frameworks, automatic software maintenance, and native integration with SageMaker features (S3, ECR, training jobs, endpoints).
- P5.4xl vs. full P5.48xlarge: The
.4xlvariant offers a single H100 GPU in a more cost-accessible form factor compared to the flagshipp5.48xlarge(8x H100), making it suitable for single-GPU experimentation and fine-tuning without the cost of a full multi-GPU node.
When to Prefer It
- Fine-tuning large language models interactively: When you need to iteratively adjust hyperparameters, inspect activations, or run LoRA/QLoRA fine-tuning on models like Llama or Mistral and want immediate feedback in a notebook environment.
- Prototyping diffusion model pipelines: When building or customizing image, video, or audio generation pipelines (e.g., Stable Diffusion variants) that require high GPU memory bandwidth and tensor compute for rapid experimentation.
- HPC and scientific computing exploration: When running exploratory pharmaceutical discovery, seismic analysis, or financial modeling workloads that benefit from H100 compute but don't yet warrant a full distributed cluster setup.
- Single-GPU model benchmarking: When you want to benchmark model performance, measure throughput, or validate quantization strategies on H100 hardware before scaling to multi-GPU training jobs.
- Cost-sensitive generative AI development: When your team needs H100-class performance but wants to avoid the cost of an 8-GPU
p5.48xlarge, and a single H100 in the.4xlform factor is sufficient for the task. - Integrated development workflows: When your workflow benefits from combining interactive notebook exploration with SageMaker's managed features (e.g., launching training jobs, deploying endpoints, accessing S3) from a single, unified environment.
Availability
- Status: Generally Available (GA) as of May 27, 2026.
- Supported regions: US East (N. Virginia), US East (Ohio), US West (Oregon), Asia Pacific (Mumbai), Asia Pacific (Tokyo), Asia Pacific (Jakarta), and South America (São Paulo).
- Pricing model: Billed at standard EC2 P5.4xl on-demand or reserved instance rates as applied through SageMaker Notebook Instance pricing; refer to the Amazon SageMaker Pricing page for current per-hour costs.
- Access method: Available when creating or modifying a SageMaker Notebook Instance by selecting the
ml.p5.4xlargeinstance type; also accessible via JupyterLab and Code Editor applications in SageMaker Studio. - Limitations: Availability is subject to regional EC2 P5 capacity; not all AWS regions are supported at launch. Standard SageMaker Notebook Instance constraints apply (single-instance, single EBS volume; data outside
/home/ec2-user/SageMakeris not persisted during maintenance).
Related Resources
- https://aws.amazon.com/ec2/instance-types/p5/
- 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
