Linux accelerated computing instances - Amazon Elastic Compute Cloud
GPU instancesVideo transcoding instancesInstances with AWS InferentiaInstances with Habana acceleratorsFPGA instancesHardware specificationsInstance performanceNetwork performanceSSD I/O performanceInstance featuresRelease notes

Linux accelerated computing instances

Accelerated computing instances use hardware accelerators, or co-processors, to perform some functions, such as floating point number calculations, graphics processing, or data pattern matching, more efficiently than is possible in software running on CPUs. These instances enable more parallelism for higher throughput on compute-intensive workloads.

If you require high processing capability, you'll benefit from using accelerated computing instances, which provide access to hardware-based compute accelerators such as Graphics Processing Units (GPUs), Field Programmable Gate Arrays (FPGAs), or AWS Inferentia.

Contents

GPU instances

GPU-based instances provide access to NVIDIA GPUs with thousands of compute cores. You can use these instances to accelerate scientific, engineering, and rendering applications by leveraging the CUDA or Open Computing Language (OpenCL) parallel computing frameworks. You can also use them for graphics applications, including game streaming, 3-D application streaming, and other graphics workloads.

G5 instances

G5 instances use NVIDIA A10G GPUs and provide high performance for graphics-intensive applications such as remote workstations, video rendering, and cloud gaming, and deep learning models for applications such as natural language processing, computer vision, and recommendation engines. These instances feature up to 8 NVIDIA A10G GPUs, second generation AMD EPY processors, up to 100 Gbps of network bandwidth, and up to 7.6 TB of local NVMe SSD storage.

For more information, see Amazon EC2 G5 Instances.

G5g instances

G5g instances use NVIDIA T4G GPUs and provide high performance for graphics-intensive applications such as game streaming and rendering that leverage industry-standard APIs, such as OpenGL and Vulkan. These instances are also suitable for running deep learning models for applications such as natural language processing and computer vision. These instances feature up to 2 NVIDIA T4G Tensor Core GPUs, AWS Graviton2 processors, and up to 25 Gbps of network bandwidth.

For more information, see Amazon EC2 G5g Instances.

G4ad and G4dn instances

G4ad instances use AMD Radeon Pro V520 GPUs and 2nd generation AMD EPYC processors, and are well-suited for graphics applications such as remote graphics workstations, game streaming, and rendering that leverage industry-standard APIs such as OpenGL, DirectX, and Vulkan. They provide up to 4 AMD Radeon Pro V520 GPUs, 64 vCPUs, 25 Gbps networking, and 2.4 TB local NVMe-based SSD storage.

G4dn instances use NVIDIA Tesla GPUs and provide a cost-effective, high-performance platform for general purpose GPU computing using the CUDA or machine learning frameworks along with graphics applications using DirectX or OpenGL. These instances provide high- bandwidth networking, powerful half and single-precision floating-point capabilities, along with INT8 and INT4 precisions. Each GPU has 16 GiB of GDDR6 memory, making G4dn instances well-suited for machine learning inference, video transcoding, and graphics applications like remote graphics workstations and game streaming in the cloud.

For more information, see Amazon EC2 G4 Instances.

G4dn instances support NVIDIA GRID Virtual Workstation. For more information, see NVIDIA Marketplace offerings.

G3 instances

These instances use NVIDIA Tesla M60 GPUs and provide a cost-effective, high-performance platform for graphics applications using DirectX or OpenGL. G3 instances also provide NVIDIA GRID Virtual Workstation features, such as support for four monitors with resolutions up to 4096x2160, and NVIDIA GRID Virtual Applications. G3 instances are well-suited for applications such as 3D visualizations, graphics-intensive remote workstations, 3D rendering, video encoding, virtual reality, and other server-side graphics workloads requiring massively parallel processing power.

For more information, see Amazon EC2 G3 Instances.

G3 instances support NVIDIA GRID Virtual Workstation and NVIDIA GRID Virtual Applications. To activate either of these features, see Activate NVIDIA GRID Virtual Applications.

G2 instances

These instances use NVIDIA GRID K520 GPUs and provide a cost-effective, high-performance platform for graphics applications using DirectX or OpenGL. NVIDIA GRID GPUs also support NVIDIA’s fast capture and encode API operations. Example applications include video creation services, 3D visualizations, streaming graphics-intensive applications, and other server-side graphics workloads.

P4d instances

These instances use NVIDIA A100 GPUs and provide a high-performance platform for machine learning and HPC workloads. P4d instances offer 400 Gbps of aggregate network bandwidth throughput and support, Elastic Fabric Adapter (EFA). They are the first EC2 instances to provide multiple network cards.

For more information, see Amazon EC2 P4d Instances.

P4d instances support NVIDIA NVSwitch GPU interconnect and NVIDIA GPUDirect RDMA.

P3 instances

These instances use NVIDIA Tesla V100 GPUs and are designed for general purpose GPU computing using the CUDA or OpenCL programming models or through a machine learning framework. P3 instances provide high-bandwidth networking, powerful half, single, and double-precision floating-point capabilities, and up to 32 GiB of memory per GPU, which makes them ideal for deep learning, computational fluid dynamics, computational finance, seismic analysis, molecular modeling, genomics, rendering, and other server-side GPU compute workloads. Tesla V100 GPUs do not support graphics mode.

For more information, see Amazon EC2 P3 Instances.

P3 instances support NVIDIA NVLink peer to peer transfers. For more information, see NVIDIA NVLink.

P2 instances

P2 instances use NVIDIA Tesla K80 GPUs and are designed for general purpose GPU computing using the CUDA or OpenCL programming models. P2 instances provide high-bandwidth networking, powerful single and double precision floating-point capabilities, and 12 GiB of memory per GPU, which makes them ideal for deep learning, graph databases, high-performance databases, computational fluid dynamics, computational finance, seismic analysis, molecular modeling, genomics, rendering, and other server-side GPU compute workloads.

P2 instances support NVIDIA GPUDirect peer to peer transfers. For more information, see NVIDIA GPUDirect.

Video transcoding instances

These instances are designed to accelerate video transcoding workloads, such as live broadcast, video conferencing, and just-in-time transcoding.

VT1 instances

VT1 instances feature Xilinx Alveo U30 media accelerators and are designed for live video transcoding workloads. These instances offer up to 8 Xilinx Alveo U30 acceleration cards, provide up to 192 GB of system memory, and up to 25 Gbps of network bandwidth. VT1 instances feature H.264/AVC and H.265/HEVC codecs and support up to 4K UHD resolutions for multi-stream video transcoding.

There are a variety of ways that you can get started:

  • Launch a VT1 instance using the Xilinx U30 AMIs on AWS Marketplace.

  • Launch a VT1 instance using your own AMI and install the Xilinx U30 drivers and Xilinx Video SDK.

  • Launch a container instance using a VT1 instance and an Amazon ECS-optimized AMI.

  • Create an Amazon EKS cluster with nodes running VT1 instances.

For more information, see Amazon EC2 VT1 Instances.

Instances with AWS Inferentia

These instances are designed to accelerate machine learning using AWS Inferentia, a custom AI/ML chip from Amazon that provides high performance and low latency machine learning inference. These instances are optimized for deploying deep learning (DL) models for applications, such as natural language processing, object detection and classification, content personalization and filtering, and speech recognition.

There are a variety of ways that you can get started:

  • Use SageMaker, a fully-managed service that is the easiest way to get started with machine learning models. For more information, see Compile and deploy a TensorFlow model on Inf1 using Sagemaker Neo.

  • Launch an Inf1 instance using the Deep Learning AMI. For more information, see AWS Inferentia with DLAMI in the AWS Deep Learning AMI Developer Guide.

  • Launch an Inf1 instance using your own AMI and install the AWS Neuron SDK, which enables you to compile, run, and profile deep learning models for AWS Inferentia.

  • Launch a container instance using an Inf1 instance and an Amazon ECS-optimized AMI. For more information, see Amazon Linux 2 (Inferentia) AMIs in the Amazon Elastic Container Service Developer Guide.

  • Create an Amazon EKS cluster with nodes running Inf1 instances. For more information, see Inferentia support in the Amazon EKS User Guide.

For more information, see Machine Learning on AWS.

Inf1 instances

Inf1 instances use AWS Inferentia machine learning inference chips. Inferentia was developed to enable highly cost-effective low latency inference performance at any scale.

For more information, see Amazon EC2 Inf1 Instances.

Instances with Habana accelerators

These instances are designed to accelerate deep learning model (DL) training workloads. They use accelerators from Habana Labs, an Intel company. These instances are optimized for DL models for applications such as image recognition, object detection and classification, and recommendation systems.

For more information, see Machine Learning on AWS.

DL1 instances

DL1 instances use Habana Gaudi accelerators. They offer up to 400 Gbps of aggregate network bandwidth, along with 32 GB of high bandwidth memory (HBM) per accelerator. DL1 instances are designed to provide high performance and cost efficiency for training deep learning models.

There are a variety of ways that you can get started:

For more information, see Amazon EC2 DL1 Instances.

FPGA instances

FPGA-based instances provide access to large FPGAs with millions of parallel system logic cells. You can use FPGA-based accelerated computing instances to accelerate workloads such as genomics, financial analysis, real-time video processing, big data analysis, and security workloads by leveraging custom hardware accelerations. You can develop these accelerations using hardware description languages such as Verilog or VHDL, or by using higher-level languages such as OpenCL parallel computing frameworks. You can either develop your own hardware acceleration code or purchase hardware accelerations through the AWS Marketplace.

The FPGA Developer AMI provides the tools for developing, testing, and building AFIs. You can use the FPGA Developer AMI on any EC2 instance with at least 32 GB of system memory (for example, C5, M4, and R4 instances).

For more information, see the documentation for the AWS FPGA Hardware Development Kit.

F1 instances

F1 instances use Xilinx UltraScale+ VU9P FPGAs and are designed to accelerate computationally intensive algorithms, such as data-flow or highly parallel operations not suited to general purpose CPUs. Each FPGA in an F1 instance contains approximately 2.5 million logic elements and approximately 6,800 Digital Signal Processing (DSP) engines, along with 64 GiB of local DDR ECC protected memory, connected to the instance by a dedicated PCIe Gen3 x16 connection. F1 instances provide local NVMe SSD volumes.

Developers can use the FPGA Developer AMI and AWS Hardware Developer Kit to create custom hardware accelerations for use on F1 instances. The FPGA Developer AMI includes development tools for full-cycle FPGA development in the cloud. Using these tools, developers can create and share Amazon FPGA Images (AFIs) that can be loaded onto the FPGA of an F1 instance.

For more information, see Amazon EC2 F1 Instances.

Hardware specifications

The following is a summary of the hardware specifications for accelerated computing instances.

Instance type Default vCPUs Memory (GiB) Accelerators
dl1.24xlarge 96 768 8
f1.2xlarge 8 122 1
f1.4xlarge 16 244 2
f1.16xlarge 64 976 8
g2.2xlarge 8 15 1
g2.8xlarge 32 60 4
g3s.xlarge 4 30.5 1
g3.4xlarge 16 122 1
g3.8xlarge 32 244 2
g3.16xlarge 64 488 4
g4ad.xlarge 4 16 1
g4ad.2xlarge 8 32 1
g4ad.4xlarge 16 64 1
g4ad.8xlarge 32 128 2
g4ad.16xlarge 64 256 4
g4dn.xlarge 4 16 1
g4dn.2xlarge 8 32 1
g4dn.4xlarge 16 64 1
g4dn.8xlarge 32 128 1
g4dn.12xlarge 48 192 4
g4dn.16xlarge 64 256 1
g4dn.metal 96 384 8
g5.xlarge 4 16 1
g5.2xlarge 8 32 1
g5.4xlarge 16 64 1
g5.8xlarge 32 128 1
g5.12xlarge 48 192 4
g5.16xlarge 64 256 1
g5.24xlarge 96 384 4
g5.48xlarge 192 768 8
g5g.xlarge 4 8 1
g5g.2xlarge 8 16 1
g5g.4xlarge 16 32 1
g5g.8xlarge 32 64 1
g5g.16xlarge 64 128 2
inf1.xlarge 4 8 1
inf1.2xlarge 8 16 1
inf1.6xlarge 24 48 4
inf1.24xlarge 96 192 16
p2.xlarge 4 61 1
p2.8xlarge 32 488 8
p2.16xlarge 64 732 16
p3.2xlarge 8 61 1
p3.8xlarge 32 244 4
p3.16xlarge 64 488 8
p3dn.24xlarge 96 768 8
p4d.24xlarge 96 1,152 8
vt1.3xlarge 12 24 2
vt1.6xlarge 24 48 4
vt1.24xlarge 96 192 16

For more information about the hardware specifications for each Amazon EC2 instance type, see Amazon EC2 Instance Types.

For more information about specifying CPU options, see Optimize CPU options.

Instance performance

There are several GPU setting optimizations that you can perform to achieve the best performance on your instances. For more information, see Optimize GPU settings.

EBS-optimized instances enable you to get consistently high performance for your EBS volumes by eliminating contention between Amazon EBS I/O and other network traffic from your instance. Some accelerated computing instances are EBS-optimized by default at no additional cost. For more information, see Amazon EBS–optimized instances.

Some accelerated computing instance types provide the ability to control processor C-states and P-states on Linux. C-states control the sleep levels that a core can enter when it is inactive, while P-states control the desired performance (in CPU frequency) from a core. For more information, see Processor state control for your EC2 instance.

Network performance

You can enable enhanced networking on supported instance types to provide lower latencies, lower network jitter, and higher packet-per-second (PPS) performance. Most applications do not consistently need a high level of network performance, but can benefit from access to increased bandwidth when they send or receive data. For more information, see Enhanced networking on Linux.

The following is a summary of network performance for accelerated computing instances that support enhanced networking.

Instance type Network performance Enhanced networking
f1.4xlarge and smaller | g3.4xlarge | g3s.xlarge | g4ad.4xlarge and smaller | g5.2xlarge and smaller | g5g.4xlarge and smaller | p3.2xlarge Up to 10 Gbps † ENA
g3.8xlarge | p2.8xlarge | p3.8xlarge 10 Gbps ENA
g5g.8xlarge 12 Gbps ENA
g4ad.8xlarge 15 Gbps ENA
g4dn.4xlarge and smaller | g5.4xlarge | inf1.2xlarge and smaller | vt1.3xlarge Up to 25 Gbps † ENA
f1.16xlarge | g3.16xlarge | g4ad.16xlarge | g5.8xlarge | g5.16xlarge | g5g.16xlarge | inf1.6xlarge | p2.16xlarge | p3.16xlarge | vt1.6xlarge 25 Gbps ENA
g5.12xlarge 40 Gbps ENA
g4dn.8xlarge | g4dn.12xlarge | g4dn.16xlarge | g5.24xlarge 50 Gbps ENA
g4dn.metal | g5.48xlarge | inf1.24xlarge | p3dn.24xlarge | vt1.24xlarge 100 Gbps ENA
dl1.24xlarge | p4d.24xlarge 4x100 Gbps ENA

† These instances have a baseline bandwidth and can use a network I/O credit mechanism to burst beyond their baseline bandwidth on a best effort basis. For more information, see instance network bandwidth.

Instance type Baseline bandwidth (Gbps) Burst bandwidth (Gbps)
f1.2xlarge 2.5 10
f1.4xlarge 5 10
g3.4xlarge 5 10
g3s.xlarge 1.25 10
g4ad.xlarge 2 10
g4ad.2xlarge 4.167 10
g4ad.4xlarge 8.333 10
g4dn.xlarge 5 25
g4dn.2xlarge 10 25
g4dn.4xlarge 20 25
g5.xlarge 2.5 10
g5.2xlarge 5 10
g5.4xlarge 10 25
g5g.xlarge 1.25 10
g5g.2xlarge 2.5 10
g5g.4xlarge 5 10
p3.2xlarge 2.5 10
vt1.3xlarge 12.5 25

SSD I/O performance

If you use a Linux AMI with kernel version 4.4 or later and use all the SSD-based instance store volumes available to your instance, you get the IOPS (4,096 byte block size) performance listed in the following table (at queue depth saturation). Otherwise, you get lower IOPS performance.

Instance Size 100% Random Read IOPS Write IOPS
g4ad.xlarge * 10,417 8,333
g4ad.2xlarge * 20,833 16,667
g4ad.4xlarge * 41,667 33,333
g4ad.8xlarge 83,333 66,667
g4ad.16xlarge 166,667 133,333
g5.xlarge * 40,625 20,313
g5.2xlarge * 40,625 20,313
g5.4xlarge * 125,000 62,500
g5.8xlarge 250,000 125,000
g5.12xlarge 312,500 156,250
g5.16xlarge 250,000 125,000
g5.24xlarge 312,500 156,250
g5.48xlarge 625,000 312,500

* For these instances, you can get up to the specified performance.

As you fill the SSD-based instance store volumes for your instance, the number of write IOPS that you can achieve decreases. This is due to the extra work the SSD controller must do to find available space, rewrite existing data, and erase unused space so that it can be rewritten. This process of garbage collection results in internal write amplification to the SSD, expressed as the ratio of SSD write operations to user write operations. This decrease in performance is even larger if the write operations are not in multiples of 4,096 bytes or not aligned to a 4,096-byte boundary. If you write a smaller amount of bytes or bytes that are not aligned, the SSD controller must read the surrounding data and store the result in a new location. This pattern results in significantly increased write amplification, increased latency, and dramatically reduced I/O performance.

SSD controllers can use several strategies to reduce the impact of write amplification. One such strategy is to reserve space in the SSD instance storage so that the controller can more efficiently manage the space available for write operations. This is called over-provisioning. The SSD-based instance store volumes provided to an instance don't have any space reserved for over-provisioning. To reduce write amplification, we recommend that you leave 10% of the volume unpartitioned so that the SSD controller can use it for over-provisioning. This decreases the storage that you can use, but increases performance even if the disk is close to full capacity.

For instance store volumes that support TRIM, you can use the TRIM command to notify the SSD controller whenever you no longer need data that you've written. This provides the controller with more free space, which can reduce write amplification and increase performance. For more information, see Instance store volume TRIM support.

Instance features

The following is a summary of features for accelerated computing instances.

EBS only NVMe EBS Instance store Placement group

DL1

No

Yes

NVMe *

Yes

F1

No

No

NVMe *

Yes

G2

No

No

SSD

Yes

G3

Yes

No

No

Yes

G4ad

No

Yes

NVMe *

Yes

G4dn

No

Yes

NVMe *

Yes
G5 No Yes NVMe * Yes
G5g Yes Yes No Yes

Inf1

Yes

No

No

Yes

P2

Yes

No

No

Yes

P3

 24xlarge: No

All other sizes: Yes

 24xlarge: Yes

All other sizes: No

 24xlarge: NVMe *

Yes

P4d

No

Yes

NVMe *

Yes

VT1

Yes

Yes

No

Yes

* The root device volume must be an Amazon EBS volume.

For more information, see the following:

Release notes

  • You must launch the instance using an HVM AMI.

  • Instances built on the Nitro System have the following requirements:

    The following Linux AMIs meet these requirements:

    • Amazon Linux 2

    • Amazon Linux AMI 2018.03

    • Ubuntu 14.04 (with linux-aws kernel) or later

    • Red Hat Enterprise Linux 7.4 or later

    • SUSE Linux Enterprise Server 12 SP2 or later

    • CentOS 7.4.1708 or later

    • FreeBSD 11.1 or later

    • Debian GNU/Linux 9 or later

  • GPU-based instances can't access the GPU unless the NVIDIA drivers are installed. For more information, see Install NVIDIA drivers on Linux instances.

  • Launching a bare metal instance boots the underlying server, which includes verifying all hardware and firmware components. This means that it can take 20 minutes from the time the instance enters the running state until it becomes available over the network.

  • To attach or detach EBS volumes or secondary network interfaces from a bare metal instance requires PCIe native hotplug support. Amazon Linux 2 and the latest versions of the Amazon Linux AMI support PCIe native hotplug, but earlier versions do not. You must enable the following Linux kernel configuration options:

    CONFIG_HOTPLUG_PCI_PCIE=y
CONFIG_PCIEASPM=y

  • Bare metal instances use a PCI-based serial device rather than an I/O port-based serial device. The upstream Linux kernel and the latest Amazon Linux AMIs support this device. Bare metal instances also provide an ACPI SPCR table to enable the system to automatically use the PCI-based serial device. The latest Windows AMIs automatically use the PCI-based serial device.

  • There is a limit of 100 AFIs per Region.

  • There is a limit on the total number of instances that you can launch in a Region, and there are additional limits on some instance types. For more information, see How many instances can I run in Amazon EC2? in the Amazon EC2 FAQ.