GPU Instances

Oracle Cloud Infrastructure (OCI) Compute provides industry-leading scalability and performance for bare metal and virtual machine (VM) instances powered by NVIDIA and AMD GPUs for mainstream graphics, AI inference, AI training, digital twins, and HPC.

Read our update on agentic AI
Twelve Labs trains AI video models with Oracle AI infrastructure (1:51)

Read the latest OCI announcements

Discover how Oracle AI infrastructure and AMD Instinct GPUs provide a high performance, scalable, and cost-effective AI/ML foundation.

Why use OCI for GPU instances?

Scalability

131,072

Maximum number of GPUs in an OCI Supercluster1

Performance

3,200

Up to 3,200 Gb/sec of RDMA cluster network bandwidth2

Value

220%

GPUs for other CSPs can be up to 220% more expensive3

Choice

VM/BM

Rightsizing with VM and performance with bare metal instances

1. OCI Supercluster scales up to 131,072 NVIDIA Blackwell B200 GPUs; more than 100,000 NVIDIA Blackwell B200 GPUs in NVIDIA Grace Blackwell GB200 Superchips; 65,536 NVIDIA H200 Tensor Core GPUs; 32,768 NVIDIA A100 Tensor Core GPUs; 16,384 NVIDIA H100 Tensor Core GPUs; and 16,384 AMD MI300X GPUs.

2. For bare metal instances with NVIDIA B200, H200, and H100 GPUs and AMD Instinct MI300X accelerators.

3. Based on on-demand pricing as of June 5, 2024.

Sign up for the new try-before-you-buy program for OCI Compute with AMD Instinct MI300X GPUs.

GPU instances—key features

OCI is the only major cloud provider to offer bare metal instances with NVIDIA and AMD GPUs for high performance that’s free of virtualization overhead. For checkpointing during AI training, our instances provide the most local storage per node (61.4 TB with H100 GPUs).

High performance NVIDIA and AMD GPUs

NVIDIA Tensor Core GPUs

OCI offers the highest value and performance for bare metal and virtual machine compute instances powered by NVIDIA Blackwell GPUs, H200 Tensor Core GPUs, H100 Tensor Core GPUs, L40S GPUs, A100 Tensor Core GPUs, A10 Tensor Core GPUs, and older-generation NVIDIA GPUs.

NVIDIA superchips

OCI offers the NVIDIA GB200 Grace Blackwell Superchip in superclusters that scale to more than 100,000 GPUs.

AMD Instinct accelerators

OCI offers AMD Instinct MI300X GPUs with 192 GB of memory at a competitive price of $6 per GPU-hour.

High performance cluster networking

Oracle’s ultralow-latency cluster networking, based on remote direct memory access (RDMA), provides microsecond-level latency.

High performance NVIDIA GPU instances
GPT-3 175B training: time to train versus number of NVIDIA H100 GPUs deployed in OCI Supercluster (0:55). Source: OCI performance for MLPerf v4.1 training

Deploy on VMs, bare metal instances, and Kubernetes clusters

VM instances

For VMs, choose from NVIDIA’s Hopper, Ampere, and older GPU architectures with one to four cores, 16 to 64 GB of GPU memory per VM, and up to 48 Gb/sec of network bandwidth.

Bare metal instances

Use OCI Supercluster with bare metal instances that include AMD Instinct GPUs, NVIDIA Blackwell GPUs or Superchips, NVIDIA Hopper GPUs or Superchips, and NVIDIA Ampere GPUs.

Kubernetes orchestration

Take advantage of managed Kubernetes, service mesh, and container registry to orchestrate AI and machine learning (ML) training and inference with containers.

Choose from a variety of VM and bare metal compute instances
Comparing the performance of NVIDIA V100 and A10 GPUs

Access readily available software

Access software and disk images

Oracle Cloud Marketplace provides software and disk images for data science, analytics, artificial intelligence (AI), and machine learning (ML) models so customers can quickly gain insight from their data.

NVIDIA AI Enterprise

Get access to NVIDIA AI Enterprise, an end-to-end software platform for data science and production AI, including generative AI, computer vision, and speech AI.

NVIDIA DGX Cloud

NVIDIA DGX Cloud on OCI is an AI-training-as-a-service platform, offering a serverless experience for developers that’s optimized for generative AI.

NVIDIA GPU Cloud Machine Image

Use NVIDIA GPU Cloud Machine Image for hundreds of GPU-optimized applications for machine learning, deep learning, and high performance computing covering a wide range of industries and workloads.

NVIDIA RTX Virtual Workstation

Deliver powerful workstation performance wherever employees need it by running NVIDIA RTX Virtual Workstation on Oracle Cloud.

Control your AI computing environment and data

Distributed cloud

When combined with GPU compute, OCI’s distributed cloud helps organizations run AI and cloud services where and how they’re needed.

Sovereign cloud

Support data residency within a region or country, including the EU, the US, the UK, and Australia.

OCI Dedicated Region

Deploy a complete cloud region in your data center with OCI Dedicated Region to retain full control of your data and applications.

Oracle Alloy

Become a partner for Oracle Alloy and deliver your cloud services to address specific market needs.

Microservices and containers

Container registry

Developers building applications using containers leverage a highly available, Oracle-managed private container registry service for storing and sharing container images. Push or pull Docker images to and from the registry using the Docker V2 API and the standard Docker command line interface (CLI). Images can be pulled directly into a Kubernetes deployment.

Oracle Functions

Functions as a service (FaaS) lets developers run serverless applications that integrate with Oracle Cloud Infrastructure, Oracle Cloud Applications, and third-party services. Gain developer efficiency along with the community of the open source Fn Project.

GPU instances—use cases

AI infrastructure for deep learning training and inferencing

Train AI models using OCI Data Science, bare metal instances, cluster networking based on RDMA, and NVIDIA GPUs.


AI training and inferencing This diagram describes two stages of deep learning model development: model training and model inferencing. In model training on the left, the untrained neural network is input to a training algorithm enabled by OCI Data Science, bare metal compute, local storage, and cluster networking. The output of the training algorithm is a trained model with a new capability. The model inferencing step is described on the right. Consider a trained model such as DALL-E 2, which can take text inputs and generate images. A text input is fed into the trained model, and an image output from the model is provided.

Virtual desktop infrastructure (VDI)

OCI Compute powered by NVIDIA GPUs provide consistent high performance for VDI.


Virtual desktop infrastructure Virtual desktop infrastructure

CFD and high performance computing using GPU instances

OCI enables computer-aided engineering and computational fluid dynamics for fast predictions of the aerodynamic properties of objects.


CFD and high performance computing using GPU instances CFD and high performance computing using GPU instances

GPU instances—customers

Explore more customer stories

Get started with GPU instances

Try Oracle AI and get a 30-day trial

Oracle offers a free pricing tier for most AI services as well as a free trial account with US$300 in credits to try additional cloud services. AI services are a collection of offerings, including generative AI, with prebuilt machine learning models that make it easier for developers to apply AI to applications and business operations.

  • Which Oracle AI and ML services offer a free pricing tier?

    • OCI Speech
    • OCI Language
    • OCI Vision
    • OCI Document Understanding
    • Machine Learning in Oracle Database
    • OCI Data Labeling

    You also only have to pay compute and storage charges for OCI Data Science.

Leverage GPUs today

Learn how Oracle helps customers leverage NVIDIA and AMD GPUs for a variety of AI use cases.

  • What can you do with GPU instances?

    • Host LLMs with NVIDIA and AMD GPUs
    • Run distributed multinode training with NVIDIA GPUs
    • Automate tasks with LLMs and retrieval-augmented generation
    • Scale NVIDIA NIM inference

See how much you can save with OCI

Oracle Cloud pricing is simple, with consistent low pricing worldwide, supporting a wide range of use cases. To estimate your low rate, check out the cost estimator and configure the services to suit your needs.

Experience the difference

  • 1/4 the outbound bandwidth costs
  • 3X the compute price-performance
  • Same low price in every region
  • Low pricing without long term commitments

Access GPU and AI experts

Get help building your next GPU solution or deploying your AI workload on OCI AI infrastructure.

  • They can answer questions such as

    • How do I get started with Oracle Cloud?
    • What kinds of AI workloads can I run on OCI?
    • What types of AI services does OCI offer?