Allszn's picture
Add launch kit overview
e193183 verified
|
Raw History Blame Contribute Delete
4.62 kB
metadata
pretty_name: AWS Steam GPU Launch Kit
license: apache-2.0
tags:
  - aws
  - infrastructure-as-code
  - cloudformation
  - gaming
  - windows
  - amazon-dcv
  - systems-manager

AWS Steam GPU Launch Kit

Deploy kit for a per-player Windows GPU host on AWS: NVIDIA gaming driver, Amazon DCV browser streaming, a Steam Offline Mode library on its own EBS volume, and IAM-scoped self-troubleshooting with a self-healing diagnostics loop.

This repo is the versioned artifact registry and the execution layer. Artifacts are fetched with huggingface_hub, so a runner that has only the AWS CLI and a token can deploy without a git checkout.

Read these three constraints first

  1. One Windows instance = one player. Amazon DCV's Windows server supports console sessions only — virtual sessions and GPU sharing are Linux-only features. Scale by adding instances, never by adding users to one box.
  2. Steam credential sharing is outside Valve's terms. Offline Mode is legitimate for your own second machine. For other people, use Steam Families (up to six members, one play slot per owned copy, shared games playable offline).
  3. Browser streaming is the convenience tier. The DCV web client is TCP-only and capped at 2 x 1920x1080. The native client gets QUIC over UDP 8443 and up to 4K.

Layout

Path Purpose
cloudformation/steam-gpu-host.yaml IAM role, security group, encrypted game volume, instance, Elastic IP, log group, status alarm
scripts/bootstrap.ps1 Installs the NVIDIA gaming driver, registers the vGaming marketplace key, installs the DCV server and the CloudWatch agent
scripts/diagnose.ps1 Emits a one-line JSON health verdict (gpu-missing, gpu-unlicensed, dcv-port-closed, ...)
scripts/validate.py Static validation, secret scanning, IAM surface inventory. No AWS credentials required.
scripts/hf_job.py Runs inside a Hugging Face Job: fetches this repo, then validates or deploys
ssm/SteamGpu-Diagnose.json Systems Manager document wrapping the diagnostic
iam/instance-role-inline-policy.json Least-privilege policy for the gaming host
iam/operator-role-policy.json Your debugging identity, with tag-scoped mutations
launch.sh One-command deploy: validate, deploy the stack, push the bootstrap over SSM
launch.py Driver with validate / plan / deploy / job modes

Validate without touching AWS

python -m pip install pyyaml
python scripts/validate.py --root .

Deploy from your own machine

export AWS_PROFILE=my-profile
export AWS_REGION=us-east-1
export KEY_NAME=my-keypair
./launch.sh --instance-type g6.xlarge --volume-gib 600

--admin-cidr defaults to your current public IP. --dry-run validates and prints the plan without changing anything.

Deploy from a Hugging Face Job

Credentials arrive as Hub secrets, so no long-lived keys sit in this repo or on disk. launch.sh then runs on ephemeral Hub compute and the runner disappears afterwards.

hf jobs uv run \
  --name steam-gpu-deploy \
  --flavor cpu-basic \
  --timeout 40m \
  --with huggingface_hub --with pyyaml --with awscli --with boto3 \
  --secrets AWS_ACCESS_KEY_ID --secrets AWS_SECRET_ACCESS_KEY --secrets AWS_REGION \
  --env KEY_NAME=my-keypair \
  --env INSTANCE_TYPE=g6.xlarge \
  scripts/hf_job.py

Set MODE=validate to run the credential-free checks on Hub compute first. Confirm the flag names against hf jobs run --help in your CLI version.

Launch console

A zero-credential parameter console lives at Allszn/steam-gpu-launch-console — it generates the exact commands above from the options you pick. It stores no credentials and calls no AWS API.

Cost

Windows GPU instances land near ~$0.71/hr (g4dn.xlarge) to ~$0.99/hr (g6.xlarge) in us-east-1, derived from Linux on-demand list rates plus AWS's stated $0.046-per-vCPU-hour Windows licence component. Verify in the AWS Pricing Calculator. Stopping the instance between sessions is the single biggest saving.

Security notes

  • No secrets are stored in this repo. scripts/validate.py fails the build if an AWS key, private key block, Steam password or bearer token is ever committed.
  • Only TCP/UDP 8443 is opened, and only to the CIDR you supply. 3389 is never exposed — reach RDP through AWS-StartPortForwardingSession.
  • The instance role cannot widen its own access; mutations in the operator role are conditioned on the Project=steam-gpu tag.