Table of Contents
The 4090's 24 GB is enough for a complete voice agent — Whisper Large-v3, Llama 3.1 8B FP16, Kokoro TTS — with room for ~10 concurrent sessions.
TL;DR
RTX 4090 hosts complete voice agent (Whisper + Llama 3.1 8B + Kokoro) using ~22 GB. ~10 concurrent voice sessions. ~700ms end-to-end latency. £289/mo.
VRAM math
| Component | VRAM |
|---|---|
| Whisper Large-v3 INT8 | 3 GB |
| Llama 3.1 8B FP16 | 16 GB |
| Kokoro TTS | 1 GB |
| KV cache @ 4K, 10 streams | ~2 GB |
| Total | ~22 GB |
Performance
- End-to-end latency: ~700 ms (vs ~480 ms on 5090)
- Concurrent voice sessions: ~10 (vs ~16 on 5090)
Verdict
4090 is a credible voice agent host at £289/mo. The 5090 at £399/mo handles ~60% more concurrency for ~30% more cost — usually worth it for production voice.
Bottom line
For 8-10 concurrent voice agents, 4090 is the budget pick. See voice agent deployment.
