Twelve and a half hours of audio processed in sixty minutes. That kind of throughput turns the RTX 3090 from a transcription tool into a transcription factory. We ran the full Whisper Large-v3 benchmark on GigaGPU to quantify exactly what this Ampere powerhouse delivers for speech recognition workloads.
Benchmark Results
| Metric | Value |
|---|---|
| Real-Time Factor (lower = faster) | 0.08 |
| Processing speed | 12.5x real-time |
| Audio hours processed per GPU-hour | 12.5 |
| Precision | FP16 |
| Performance rating | Very Good |
Benchmark conditions: FP16 inference, single-stream processing, 16kHz input audio, English language. faster-whisper backend with CTranslate2 optimisation.
A Model That Barely Touches 24 GB
| Component | VRAM |
|---|---|
| Model weights (FP16) | 3.1 GB |
| Audio buffer + runtime | ~0.5 GB |
| Total RTX 3090 VRAM | 24 GB |
| Free headroom | ~20.9 GB |
Over 20 GB free. The 3090 is massively overprovisioned for Whisper alone, and that is the point. Run a complete voice AI stack: Whisper for transcription, a 7B LLM for response generation, and Coqui XTTS-v2 for speech synthesis — all simultaneously on one card with VRAM to spare.
Cost per Audio Hour
| Cost Metric | Value |
|---|---|
| Server cost | £0.75/hr (£149/mo) |
| Cost per audio hour | £0.060 |
| Audio hours per £1 | 16.7 |
Six pence per audio hour at 12.5x speed. For context, that means processing an entire day’s worth of call centre recordings (100+ hours) overnight for about £6. See how this stacks up against other GPUs on the benchmark comparison page.
Built for High-Volume Audio
The 3090 is the right card for organisations with serious transcription volume: legal firms processing depositions, media companies transcribing interviews, healthcare providers converting clinical notes. The 12.5x speed means backlogs disappear fast, and the VRAM surplus means you are never locked into a single workload. For even more throughput, the RTX 5090 reaches 33.3x. Full details: best GPU for Whisper.
Quick deploy:
docker run --gpus all -p 9000:9000 ghcr.io/fedirz/faster-whisper-server:latestExplore: Whisper hosting guide, all benchmarks, PaddleOCR hosting.
Deploy Whisper Large-v3 on RTX 3090
Order this exact configuration. UK datacenter, full root access.
Order RTX 3090 Server
