Home → Speech
Whisper VRAM and speed by model
Whisper is light by modern standards — even the largest tier fits in 10GB. The interesting axis is speed: the small tiers are dramatically faster per unit of audio.
Model table
| Model | Approx. VRAM | Relative speed |
|---|---|---|
| tiny | ~1GB | ~10× |
| base | ~1GB | ~7× |
| small | ~2GB | ~4× |
| medium | ~5GB | ~2× |
| large-v3 | ~10GB | 1× |
| turbo | ~6GB | ~8× |
Source: openai/whisper official README (relative speed vs large, retrieved Oct 2026).
How to read relative speed
A 10× model transcribes an hour of audio in about six minutes of compute; 1× takes about an hour. The turbo tier is the notable entry: near-large quality positioning at ~6GB and ~8× speed, which is why it has become the default for bulk transcription jobs.
Choosing for a batch of audio
- Any modern 8GB+ card runs every tier; the decision is accuracy vs throughput, not memory.
- For long-running bulk jobs, throughput per watt matters — the small/turbo tiers process far more audio per GPU-hour than large-v3.
- Speech workloads are bursty, which is why many teams run them as short rented jobs instead of keeping a GPU idle between batches.