I am running an agent server on 2cpu and 4 gb ram....
# ask-aif
fancy-florist-18763
02/25/2026, 6:56 PMI am running an agent server on 2cpu and 4 gb ram. should it be enough?
t
tall-airline-16756
02/25/2026, 6:57 PMThanks for reaching out, Palwisha. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.
tall-airline-16756
02/25/2026, 6:57 PMThanks — I’m happy to help. You’re asking whether 2 CPU / 4 GB RAM is enough for an agent server; I’ll cover the short answer, when it might be OK, and a couple of quick follow-ups so I can give a more precise recommendation.
Short answer
• 2 CPUs and 4 GB RAM is below LiveKit’s recommended starting point for most voice AI agent workloads. It may be borderline or insufficient for production voice agents, especially if you have concurrent sessions or use CPU-intensive features.
When 2 CPU / 4 GB might be acceptable
• Lightweight, low-concurrency workloads: simple transcription-only agents, no heavy real‑time audio processing, and only a few concurrent sessions.
• Development, testing, or proof-of-concept with very low traffic.
When you’ll likely need more resources
• Voice agents that use noise cancellation, real‑time VAD/TTS/STT, or other CPU-heavy media processing.
• Moderate to high concurrency (tens of concurrent sessions).
• Production deployments that must remain responsive under load.
Recommended baseline
• LiveKit recommends starting around 4 CPU cores and 8 GB RAM for most voice AI applications; that configuration can handle roughly 10–25 concurrent jobs depending on components used. Real-world tests reached near 4 cores and ~2.8 GB RAM while handling ~30 concurrent voice-to-voice sessions.
Operational note
• When you restart or scale down agent servers they stop accepting new jobs on SIGINT/SIGTERM; allow a graceful shutdown period (for voice agents, consider 10+ minutes) so in‑flight sessions can finish.
To give a tailored recommendation, could you tell me:
1. What type of agent workload is this? (e.g., simple transcription, voice‑to‑voice with TTS/STT, or LLM-driven dialog)
2. Expected concurrent sessions or jobs at peak?
3. Are you running any extra processing (noise suppression, real‑time VAD, turn detection, custom models)?
4. Is this for development/testing or production?
With those details I’ll suggest whether to stick with 2/4, upgrade, or optimize the agent configuration.
Sources: Self-hosted deployments | LiveKit Documentation
f
fancy-florist-18763
02/25/2026, 6:59 PMim using stt, llm and tts agent session configuration
t
tall-airline-16756
02/25/2026, 6:59 PMThanks — that’s helpful. You’re running an agent with STT, LLM, and TTS. I’ll be direct: 2 CPU and 4 GB RAM is likely insufficient for a production voice agent using all three components, unless you expect very low concurrency and minimal processing per session.
Why:
• STT and TTS both add real‑time audio processing overhead.
• LLM-driven sessions add CPU (and possibly memory) usage depending on whether the model runs locally or you call an external API.
• Real-world tests on similar voice-to-voice setups showed a 4‑core, 8 GB machine reaching ~3.8 CPU cores and ~2.8 GB RAM while handling dozens of concurrent sessions — that implies a 2/4 instance will become CPU‑bound quickly under normal load.
What I recommend next
1. Tell me your expected concurrency (peak number of simultaneous sessions).
2. Confirm whether your LLM is hosted locally on the agent server or called via an external API (remote).
3. Let me know whether you use extra processing like noise suppression, VAD, or turn detection.
Quick guidance depending on answers
• If concurrency is very low (1–3 concurrent sessions) and LLM calls are remote: you might be able to run on 2 CPU / 4 GB for development or small tests, but watch CPU load and latency.
• If you expect >5 concurrent sessions, or run the LLM locally, or use noise suppression/VAD: upgrade to at least 4 CPUs and 8 GB RAM as a starting point.
• For production with moderate concurrency or strict latency SLAs: consider scaling horizontally (more instances sized at 4 CPU / 8 GB) or larger single instances.
If you share concurrency and whether the LLM is local vs remote, I’ll give a concrete sizing recommendation and suggested monitoring metrics to watch (CPU, memory, and request latency).
5 Views