Welche GPU für Llama 3.3 70B lokal?
Llama 3.3 70B hat 70 Mrd. Parameter. Diese GPUs können das Modell in mindestens einer praktikablen Quantisierung fahren — sortiert nach Preis pro GB VRAM.
| GPU | beste Quant | VRAM-Bedarf | Tokens/Sek | Preis | €/GB |
|---|---|---|---|---|---|
| Apple Mac mini M4 Pro 64GB | Q4_K_M | 42.8 GB | — | — | — |
| Apple Mac Studio M3 Ultra 192GB | Q8_0 | 72.6 GB | — | — | — |
| Apple Mac Studio M3 Ultra 96GB | Q8_0 | 72.6 GB | — | — | — |
| Apple MacBook Pro M4 Max 128GB | Q8_0 | 72.6 GB | — | — | — |
| Apple MacBook Pro M4 Max 64GB | Q4_K_M | 42.8 GB | — | — | — |
| NVIDIA H100 80GB | Q8_0 | 72.6 GB | — | — | — |
| NVIDIA L40S | Q4_K_M | 42.8 GB | — | — | — |
| NVIDIA RTX 6000 Ada Generation | Q4_K_M | 42.8 GB | — | — | — |
| NVIDIA RTX A6000 | Q4_K_M | 42.8 GB | — | — | — |