A model that doesn't fit doesn't run slowly, it doesn't run. Here's every card from our bench, sorted by measured speed.
Weights: hexgrad/Kokoro-82M
| GPU | Speed | VRAM | Basis |
|---|---|---|---|
| NVIDIA L40S | 243.99 x realtime | 48GB | ✓ Measured |
| NVIDIA H100 80GB HBM3 | 222.02 x realtime | 80GB | ✓ Measured |
| NVIDIA H200 | 170.82 x realtime | 141GB | ✓ Measured |
| NVIDIA A100 40GB SXM4 | 142.03 x realtime | 40GB | Est. |
| NVIDIA A100 80GB SXM4 | 110.17 x realtime | 80GB | ✓ Measured |
| NVIDIA A10G | 101.11 x realtime | 24GB | ✓ Measured |
| NVIDIA L4 | 97.35 x realtime | 24GB | ✓ Measured |
| NVIDIA T4 | 42.88 x realtime | 16GB | ✓ Measured |
NVIDIA L40S is the fastest card we've measured running Kokoro TTS 82M, at 243.99 x realtime.
NVIDIA T4 ($2,299 MSRP) is the cheapest card on our bench that runs Kokoro TTS 82M, at 42.88 x realtime.