EC2 Windows GPU: Install NVIDIA GRID and Publish CloudWatch Metrics

A standard Windows Server AMI on g4dn or g5 may initially expose only Microsoft Basic Display Adapter, with no working nvidia-smi. After installing the driver, another gap appears: the Windows CloudWatch Agent cannot use the Linux nvidia_gpu receiver.

This guide connects both steps—installing an AWS GRID driver and publishing parsed nvidia-smi values as custom CloudWatch metrics.

Select and install the driver

GRID is appropriate for graphics, remote desktops, video encode, and WDDM workloads. Data Center/Tesla drivers are commonly used for compute/TCC workloads.

List the current files in the AWS public bucket instead of hard-coding a version from an article:

aws s3 ls s3://ec2-windows-nvidia-drivers/latest/ `
  --no-sign-request --region us-east-1

Choose the file matching the Windows release:

$ErrorActionPreference = 'Stop'
$driver = 'C:\ProgramData\Amazon\NVIDIA\grid-driver.exe'
New-Item (Split-Path $driver) -ItemType Directory -Force | Out-Null

aws s3 cp `
  's3://ec2-windows-nvidia-drivers/latest/<selected-driver>.exe' `
  $driver --no-sign-request --region us-east-1

$p = Start-Process $driver `
  -ArgumentList '-s -noreboot -noeula' -Wait -PassThru
if ($p.ExitCode -ne 0) { throw "Driver installer exit code $($p.ExitCode)" }

& "$env:ProgramFiles\NVIDIA Corporation\NVSMI\nvidia-smi.exe"

A reboot may be required. Before a remote reboot, verify that SSM, EC2Launch, or another recovery path is available.

IAM permission

The instance role needs cloudwatch:PutMetricData. The permission can be constrained to the custom namespace:

{
  "Version": "2012-10-17",
  "Statement": [{
    "Effect": "Allow",
    "Action": "cloudwatch:PutMetricData",
    "Resource": "*",
    "Condition": {"StringEquals": {"cloudwatch:namespace": "Custom/GPU"}}
  }]
}

Collector

Use IMDSv2 for region and instance dimensions, parse every GPU row, and submit multiple values in one API call per GPU:

$ErrorActionPreference = 'Stop'
$smi = "$env:ProgramFiles\NVIDIA Corporation\NVSMI\nvidia-smi.exe"

$token = Invoke-RestMethod -Method Put `
  -Uri 'http://169.254.169.254/latest/api/token' `
  -Headers @{'X-aws-ec2-metadata-token-ttl-seconds'='21600'}
$h = @{'X-aws-ec2-metadata-token'=$token}
$region = Invoke-RestMethod -Headers $h `
  -Uri 'http://169.254.169.254/latest/meta-data/placement/region'
$instanceId = Invoke-RestMethod -Headers $h `
  -Uri 'http://169.254.169.254/latest/meta-data/instance-id'
$instanceType = Invoke-RestMethod -Headers $h `
  -Uri 'http://169.254.169.254/latest/meta-data/instance-type'

$rows = & $smi `
  '--query-gpu=index,utilization.gpu,utilization.memory,memory.total,memory.used,temperature.gpu,power.draw' `
  '--format=csv,noheader,nounits'

foreach ($row in $rows) {
  $v = $row.Split(',') | ForEach-Object { $_.Trim() }
  $dims = @(
    @{Name='InstanceId';Value=$instanceId},
    @{Name='InstanceType';Value=$instanceType},
    @{Name='GPUIndex';Value=$v[0]}
  )
  $metrics = @(
    @{MetricName='GPUUtilization';Value=[double]$v[1];Unit='Percent';Dimensions=$dims},
    @{MetricName='MemoryUtilization';Value=[double]$v[2];Unit='Percent';Dimensions=$dims},
    @{MetricName='MemoryTotal';Value=[double]$v[3];Unit='Megabytes';Dimensions=$dims},
    @{MetricName='MemoryUsed';Value=[double]$v[4];Unit='Megabytes';Dimensions=$dims},
    @{MetricName='Temperature';Value=[double]$v[5];Unit='None';Dimensions=$dims},
    @{MetricName='PowerDraw';Value=[double]$v[6];Unit='None';Dimensions=$dims}
  )
  aws cloudwatch put-metric-data --namespace 'Custom/GPU' `
    --metric-data ($metrics | ConvertTo-Json -Depth 5 -Compress) `
    --region $region
}

Save it as C:\ProgramData\Amazon\GPUMetrics\collect.ps1 and schedule it once per minute:

$action = New-ScheduledTaskAction -Execute 'powershell.exe' -Argument `
  '-NoProfile -ExecutionPolicy Bypass -File "C:\ProgramData\Amazon\GPUMetrics\collect.ps1"'
$trigger = New-ScheduledTaskTrigger -Once -At (Get-Date).AddMinutes(1) `
  -RepetitionInterval (New-TimeSpan -Minutes 1)
$principal = New-ScheduledTaskPrincipal -UserId 'SYSTEM' `
  -LogonType ServiceAccount -RunLevel Highest
Register-ScheduledTask -TaskName 'GPU-Metrics-Collector' `
  -Action $action -Trigger $trigger -Principal $principal -Force

Validation and alerts

Get-ScheduledTaskInfo 'GPU-Metrics-Collector'
aws cloudwatch list-metrics --namespace 'Custom/GPU' --region $region

Useful alerts include sustained high temperature, prolonged saturation, memory exhaustion, and power unexpectedly dropping to zero. Thresholds must be based on the GPU model and workload baseline.

Common issues:

• Missing nvidia-smi: driver installation is incomplete or a reboot is pending.
• No metrics: inspect IAM, the namespace condition, and LastTaskResult.
• Multiple GPUs: parse every CSV row and include GPUIndex.
• System proxy: bypass 169.254.169.254 or both IMDS and SSM may fail.
• Custom metrics and API requests incur charges; check current regional pricing.

Summary

Treat “the driver works” and “the workload is observable” as separate acceptance criteria. Select the current matching GRID package, verify nvidia-smi, then batch custom metric data into CloudWatch for dashboards and alarms across the Windows GPU fleet.