GPU Nodes

A GPU node is a game server node that has a GPU and is enabled for rendering work. GPU nodes power everything in 5Stack that produces video:

• In-browser demo playback
• Live streaming
• Highlights and clip rendering
• Utility lineup preview clips and stills

Without a GPU node, the panel still works, matches, stats, and 2D/3D replay all run fine, but those video features are unavailable.

WSL is not supported

GPU nodes will not work under WSL (Windows Subsystem for Linux). Game streaming, highlights/clip rendering, and in-browser demo playback all require a native Linux host with direct GPU access. Use a bare-metal or virtualized Linux install instead.

How it works

A GPU node runs real CS2 instances to render demos and live observers in a GPU pod. Each render job (a demo session, a live stream, or a highlight render) boots a pod, does its work, and frees the GPU again. Lineup previews work a little differently: one pod films the whole queue (see Utility lineup previews). The GPU Nodes admin page shows the node roster with live GPU/VRAM usage, what each node is currently doing, and how many GPUs are free in the pool.

Adding a GPU node

1. Create a game server node on the machine with the GPU.
2. Install the GPU driver on that node.
3. Run setup-gpu-nodes.sh from your panel install directory.
4. Add Steam accounts to the pool on the GPU Nodes page.

Create the node

A GPU node starts out as an ordinary game server node. Create it the same way, on a native Linux host with the GPU installed.

Install the GPU driver

Every GPU node needs its vendor's driver so render pods can reach the GPU. NVIDIA also needs its container runtime. Follow the guide for your GPU:

Vendor	Status
NVIDIA	Supported
AMD	Supported on Strix Halo, experimental on other Radeon GPUs
Intel	Not yet supported, tracking issue #468

Run the setup script

From your panel install directory, run:

./setup-gpu-nodes.sh

The script will walk you through:

1. GPU vendor selection, NVIDIA, AMD, or Intel. Intel exits early with a link to the tracking issue.
2. GPU node selection, pick which cluster nodes have GPUs (multi-select). The script labels selected nodes with 5stack-game-streamer=true, plus nvidia-gpu=true on NVIDIA or 5stack-gpu-vendor=amd on AMD. Re-running the script with a different selection will add the labels to newly-selected nodes and remove them from previously-selected nodes.
3. Integrated GPUs (AMD only), pick which of those nodes should render on the GPU integrated into their CPU. Leave nodes with a discrete Radeon card unticked. See Integrated GPUs.
4. Steam credentials, username and password for a dedicated account the render pods use to download CS2. Asked only when they aren't saved yet.
5. Playback domain, the domain viewers' browsers load game streams from (default hls.<your web domain>). Asked every run, with the current value as the default.

Your selections are persisted to .5stack-env.config, and the script finishes by running update.sh to apply them. On NVIDIA, that also deploys the Kubernetes device plugin so pods can request the GPU. AMD needs no device plugin.

Steam account must NOT have 2FA / Steam Guard

The render pods log into Steam non-interactively via steamcmd to download and run CS2. Steam Guard / Steam Mobile Authenticator will break this, every Job will get stuck waiting for a code that nobody can enter.

Use a dedicated Steam account with 2FA disabled. Do not reuse a personal account.

Steam account pool

To render a demo, observe a live match or film a lineup preview, the pod has to log into CS2 with a Steam account. GPU nodes draw from a shared Steam account pool that you manage from the GPU Nodes page. Add at least as many accounts as you have GPU nodes so each node can claim one; the page shows how many accounts are available versus how many nodes are registered.

WARNING

Use dedicated Steam accounts for the pool, not your personal account. An account in use by a render job can't be used to play at the same time.

Shader baking

The first time a GPU node renders, CS2 compiles its shader cache, which can take a while. 5Stack can bake shaders ahead of time from the GPU Nodes page so the first real render isn't slowed down by compilation. Bake progress and logs are shown inline per node.

Utility lineup previews

A lineup's preview clip and stills are filmed by a render pod that joins a practice server as a player while the practice plugin directs the shots. The pod connects to the server over loopback, so the two always run on the same node:

• The practice server is only booked on a GPU node that is enabled, has the Rendering workload turned on, and has a free GPU. Until one is free, the queue waits and checks again every minute.
• The pod logs into CS2 with an account from the Steam account pool, and waits if none is free.
• The practice server respects Reserved on-demand slots, like any practice server, so previews cannot take the last slot from a scheduled match.
• Pause Highlight Renders During Active Match keeps lineup renders off a node while a match is live on it too. The render's own practice server does not count.

One pod and one practice server film the whole queue. Between lineups the pod asks the API what to film next: another lineup on the same map, a change of level onto the next map in the queue, a short wait, or stop. The practice server changes map instead of being replaced, and the GPU stays held while the pod keeps asking. A live stream waiting for a GPU gets it at the next map change, or when the queue runs dry. The pod stops after the queue has been empty for a little over a minute, or after three lineups in a row fail, so the next batch starts with a fresh client.

Each lineup is captured at 1080p and 60 fps, re-encoded with NVENC on NVIDIA or VA-API on AMD (x264 under CPU only, or when Auto can't use the GPU encoder) so a shared link plays inline in Discord, with six stills cut from the capture. Track the queue under System → Render Queue, on the Utility Lineups tab.

The practice server for a render runs as the game-server job u-<match id>, and the render pod as gs-nades-queue.

Discover needs no GPU

Discover also throws on a practice server, but nothing is filmed, so its server is not tied to a GPU node.

Image overrides

UTILITY_RENDER_STREAMER_IMAGE and UTILITY_RENDER_SERVER_IMAGE on the API override the game-streamer and game-server images for lineup renders only (the server image also applies to Discover). Without them, renders use the same images as everything else.

Monitoring

The GPU Nodes page surfaces, per node: status and region, GPU model and VRAM, live GPU/memory graphs, the current task (live stream, demo playback, or highlight render) with a snapshot preview, and shader-bake state. Pool-level counters at the top show free vs. registered GPUs and Steam pool usage.