Hardware requirements

This page matters when you host the stack yourself — on-prem cloud or a kiosk. On hosted SaaS, none of it is your problem.

The short version

Scenario	GPU
2D avatar, rendering only	RTX 3060 — minimum
Full stack on one box (recognition + retrieval + model + speech + 2D avatar)	RTX 5070 Ti Mobile — recommended
3D Unreal Engine avatar	Lower than 2D; can render on the end-user device, including phones

The RTX 5070 Ti Mobile spec is our standard kiosk recommendation for both 2D and 3D deployments. It carries speech recognition, knowledge retrieval, the language model, speech synthesis, and 2D avatar rendering simultaneously on a single machine, which is what makes a self-contained kiosk possible.

2D avatars

The RTX 3060 is the floor: it renders a 2D avatar at full frame rate with hardware video encoding, but it leaves little headroom. On a 3060 you should expect to run the language model somewhere else — a hosted API or a second box.

Above that floor the useful upgrades are, in order: VRAM (decides which model fits alongside the renderer), then encoder throughput (decides how many concurrent streams the box can push).

Requirements per concurrent conversation, not per kiosk: a screen that is idle costs nothing.

3D avatars

Unreal Engine rendering is materially cheaper than neural 2D rendering, and it does not have to happen on the server at all. Rendered client-side, a modern smartphone handles the avatar itself and the server only carries the voice pipeline. That is the cheapest way to put an avatar in a consumer mobile app.

Rendered server-side, use the same recommendation as 2D.

A single self-contained kiosk

Component	Recommended
GPU	RTX 5070 Ti Mobile (or a desktop card of equivalent class)
CPU	8+ modern cores
RAM	32 GB
Storage	250 GB NVMe — model weights and avatar assets dominate
Audio	A directional USB microphone and powered speakers. This is the single biggest driver of perceived quality in a noisy hall — budget for it properly
Display	Any; portrait 1080×1920 is the common kiosk format
Network	Wired ethernet strongly preferred over Wi-Fi

Appliance plus edge screens

When a site has several screens, one GPU host on the LAN runs the voice pipeline and the models, and each screen runs a thin edge box that renders and displays. One appliance comfortably serves 4–12 screens depending on how many people talk at once. Edge boxes need a modest GPU for decode and display, not for inference.

On-prem cloud sizing

Two VMs per customer: a small proxy host (2 vCPU / 4 GB is plenty) and a GPU host. An L4-class GPU or better is the usual starting point for the GPU host; concurrency, not raw speed, decides how far up that scale you go. We size it with you against expected peak concurrent conversations, then leave headroom.

Network

Requirement	Detail
Bandwidth	~1–1.5 Mbps per concurrent conversation (video plus audio, both directions)
Transport	WebRTC over UDP preferred; a TCP relay fallback is included for restrictive networks
Ports out	443/TCP and 443/UDP; a relay port (3478) when clients sit behind strict NAT
Latency	Users and GPU in the same region. Cross-continent placement is the most common cause of a deployment that "feels slow"

Corporate networks that block UDP outright still work through the relay fallback, at the cost of some latency. Tell us early if that is your situation — it changes how we place the relay.

Power and physical

Kiosk hardware runs hot under sustained GPU load. Ensure ventilation, and plan for unattended reboots: the stack restarts itself into a working state after power loss, with no operator present.