AWS Amazon EC2

g7.48xlarge

⚡ 8x GPU • 256 GB

Part of the G7 Accelerated Computing family — Hardware accelerators (GPUs, FPGAs) for specialized compute.

Name:gGPU Graphics & ML Inference77th Generation48xlarge48 Extra Large

On-demand, Linux

$28.5133/ hr

~$20,814.69/mo · lowest in US East (N. Virginia), us-east-1

Savings Plan, Linux

$15.39717/ hr

From, with a 3-year Compute Savings Plan, No Upfront: −46% vs on-demand rate (us-east-1)

Per unit

$0.1485/ vCPU-hr

$0.0371/GB-hr

Compute

192vCPUs

192 cores • x86_64

Memory

768GB

4.0 GB / vCPU

Network

700 Gbps

Baseline: 700 Gbps

Storage & EBS

10,000MB/s

Max: 320,000 IOPS

On Google Cloud

g7.48xlarge is an accelerated machine. Those are compared by the accelerator, not by vCPU and memory, so this page doesn't pick a match. Browse Google Cloud machines.

Technical Specifications & Limits

Hardware specs, container thresholds, and system limits

Compute & Processor

ProcessorIntel Xeon Scalable (Granite Rapids)

Architecturex86_64 (x86_64)

vCPUs / Threads192 vCPUs

Physical Cores192

Clock Speed3.9 GHz

Intel AVX / AVX2AVX & AVX2 Supported

Intel Turbo BoostYes

Bare MetalNo (Virtualized)

Performance & Benchmarks

CoreMark Score5,701,982

CoreMark / vCPU29,697.8 / core

FFmpeg Transcoding301 FPS

CoreMark per Dollar199,976 score / $

Cost per vCPU-hr$0.1485

Cost per GB-hr$0.0371

Containers & Kubernetes

EKS Max Pods758 pods

Max ENIs24

IPv4/v6 per ENI64

Total Addressable IPs1536 total IPs

CNI Prefix DelegationSupported (High Density)

Networking & Bandwidth

Network Performance700 Gbps

Baseline Bandwidth700 Gbps

Burst / Peak Bandwidth700 Gbps

Enhanced NetworkingYes

EFA SupportNo

IPv6 SupportNative Dual-Stack Supported

Storage & EBS Optimization

Storage ConfigurationNVMe SSD (7600 GB)

Local NVMe Disk7,600 GB NVMe SSD

EBS Baseline Throughput10000 MB/s

EBS Max Throughput10000 MB/s

EBS Baseline IOPS320,000

EBS Max / Burst IOPS320,000

Dedicated EBS Throughput80000 Mbps

GPU & Hardware Acceleration

GPU ModelNVIDIA RTX PRO 4500

Accelerator Count8x GPU(s)

Total GPU Memory256.0 GB

Memory Architecture—

Acceleration RuntimesCUDA, TensorRT, cuDNN, PyTorch, PyJAX

Platform & Architecture Details

Hypervisor / SystemAWS Nitro System (PCIe Offload)

Nitro EnclavesSupported (Isolated security enclaves)

Normalization Factor11.31479365

Generation StatusCurrent Generation (Recommended)

Rates by region

Live on-demand rates across all global AWS cloud regions

Lowest Price Region

$28.5133/ hr

US East (N. Virginia) (us-east-1)

Highest Price Region

$35.6563/ hr

EU (Frankfurt) (eu-central-1)

Global Fleet Average

$30.2295/ hr

Across all active regions

Regional Coverage

6regions

Available worldwide

Linux — Available in all 6 regions
Region	On-Demand Price	% vs On-Demand Rate
eu-central-1 EU (Frankfurt) Highest	$35.6563
eu-north-1 EU (Stockholm)	$30.2422
eu-south-2 Europe (Spain)	$29.9389
us-east-1 US East (N. Virginia) Lowest	$28.5133
us-east-2 US East (Ohio) Lowest	$28.5133
us-west-2 US West (Oregon) Lowest	$28.5133

Windows — Available in all 6 regions
Region	On-Demand Price	% vs On-Demand Rate
eu-central-1 EU (Frankfurt) Highest	$44.4883
eu-north-1 EU (Stockholm)	$39.0742
eu-south-2 Europe (Spain)	$38.7709
us-east-1 US East (N. Virginia) Lowest	$37.3453
us-east-2 US East (Ohio) Lowest	$37.3453
us-west-2 US West (Oregon) Lowest	$37.3453

Savings Plans: what to know

A Savings Plan commits you to a dollar amount per hour of spend for one or three years, not to an instance: usage up to that amount is billed at the plan's rates and the rest at on-demand, and unused commitment is still billed.
A Compute Savings Plan follows your usage across instance families, sizes, regions and operating systems; an EC2 Instance Savings Plan is locked to one family in one region in exchange for a lower rate (what that flexibility costs).
In US East (N. Virginia) (us-east-1), the 3-year Compute Savings Plan rate (No Upfront) for g7.48xlarge on Linux is 54% of the on-demand rate, so the instance has to run at least 54% of the hours for the commitment to cost less than on-demand.
On Windows a plan discounts only the compute part of the rate: the Windows license part costs the same as on demand, so the percentage off is smaller than on Linux.
Rates are AWS's effective hourly rate for each payment option; All Upfront pays the whole term's commitment at purchase, which is why its hourly rate is lower; Partial Upfront (at least half up front) isn't shown on this page. The Savings Plans calculator has all three and sizes a commitment for this instance or a mix of them. How AWS applies a plan each hour: how Compute Savings Plans decide what they cover. Sizing one for elastic workloads: managing Compute Savings Plans.