Chatbots that can talk back need two things on the GPU at once: an LLM to produce text and a TTS model to vocalise it. We co-hosted LLaMA 3 8B and Coqui XTTS-v2 on a single NVIDIA RTX 3090 (24 GB VRAM) running on a GigaGPU dedicated server. This benchmark shows what happens when both models share the same GPU under real concurrent load.
Models tested: LLaMA 3 8B + Coqui XTTS-v2
How Both Models Perform Together
| Component | Metric | Solo | Concurrent |
|---|---|---|---|
| LLaMA 3 8B (FP16) | Tokens/sec | 62 | 44.6 |
| Coqui XTTS-v2 | Real-time factor | 0.18 | 0.216 |
| Coqui XTTS-v2 | Synthesis speed | 5.6x | 4.6x |
All models loaded simultaneously in GPU memory. Throughput figures reflect concurrent operation with shared VRAM and compute.
Memory Pressure
| Component | VRAM |
|---|---|
| Combined model weights | 20.4 GB |
| Total RTX 3090 VRAM | 24 GB |
| Free headroom | ~3.6 GB |
Both models fit with 3.6 GB to spare. That is tighter than the Whisper pipeline because XTTS-v2 carries a larger neural vocoder. The headroom is adequate for normal inference, but if you need longer LLM context windows or plan to stream synthesis for extended passages, quantising the LLM to INT4 would free up significant room.
Cost Advantage of a Single GPU
| Cost Metric | Value |
|---|---|
| Server cost (single GPU) | £0.75/hr (£149/mo) |
| Equivalent separate GPUs | £1.50/hr |
| Savings vs separate servers | 50% |
Running LLM and TTS on one card at £149/mo eliminates the need for a separate TTS server. The LLM generates text, Coqui synthesises speech at 4.6x real-time — fast enough that spoken responses are ready well before a human listener would expect them. Full numbers at our benchmark page.
Real-World Application
The LLM + TTS pipeline matters for any product that needs a natural-sounding voice: IVR systems, AI tutors, accessibility tools, or kiosk assistants. At 44.6 tok/s generation feeding into 4.6x real-time synthesis, the 3090 handles conversational cadences without noticeable lag. The key constraint is VRAM. If your TTS use case grows to include voice cloning with longer reference samples or you want to add a Whisper front-end, you will want to step up to the RTX 5080 or 5090.
Quick deploy:
docker compose up -d # llama.cpp + xtts-streaming-server containers with --gpus allSee our LLM hosting guide, Coqui TTS hosting guide, and all benchmark results. Related benchmarks: LLaMA 3 8B on RTX 3090, Coqui XTTS-v2 on RTX 3090.
Deploy LLM + TTS Pipeline on RTX 3090
Order this exact configuration. UK datacenter, full root access.
Order RTX 3090 Server
