Mac or dedicated GPU?
Memory decides which models you can load. Bandwidth decides how fast they run. Macs and GPUs trade these off differently.
Unified memory (Macs and similar)
The CPU and GPU share one pool of memory. A 64 GB Mac can hold models no single consumer GPU can. We count about 75% of it as usable, because macOS needs the rest. Generation is usually slower than on a high-end GPU, but it is quiet and power-efficient.
VRAM (discrete GPUs)
The memory on the graphics card is very fast, so tokens/sec is high, but it is capped at the card's size. We count about 90% as usable. A model that doesn't fit spills into system RAM ("partial offload") and slows down sharply.
Popular options, cheapest first
| Device | Memory | Usable (est.) | Price | $/GB |
|---|---|---|---|---|
| Qualcomm Snapdragon 8 Gen 3 Leading Version | 12 GB unified | 9.0 GB | $0 | $0 |
| Qualcomm Snapdragon 8 Elite | 12 GB unified | 9.0 GB | $0 | $0 |
| Qualcomm Snapdragon X Elite | 16 GB unified | 12.0 GB | $0 | $0 |
| Intel Arc A770 8GB | 8 GB VRAM | 7.2 GB | $299 | $37 |
| Intel Arc A770 16GB | 16 GB VRAM | 14.4 GB | $329 | $21 |
| Mac Mini M4 | 16 GB unified | 12.0 GB | $599 | $37 |
| AMD Radeon RX 7800 XT | 16 GB VRAM | 14.4 GB | $599 | $37 |
| AMD Radeon RX 7900 XT | 20 GB VRAM | 18.0 GB | $899 | $45 |
| Mac Mini M4 32GB | 32 GB unified | 24.0 GB | $999 | $31 |
| AMD Radeon RX 7900 XTX | 24 GB VRAM | 21.6 GB | $999 | $42 |
| Mac Mini M4 Pro | 24 GB unified | 18.0 GB | $1399 | $58 |
| MacBook Pro M4 14-inch | 16 GB unified | 12.0 GB | $1599 | $100 |
| NVIDIA RTX 4090 | 24 GB VRAM | 21.6 GB | $1599 | $67 |
| NVIDIA RTX 5090 | 32 GB VRAM | 28.8 GB | $1999 | $62 |
| Mac Studio M5 Max | 36 GB unified | 27.0 GB | $2500 | $69 |
| MacBook Pro M4 Max 16-inch | 48 GB unified | 36.0 GB | $3999 | $83 |
| Mac Studio M5 Ultra | 96 GB unified | 72.0 GB | $5500 | $57 |
| NVIDIA RTX 6000 Ada | 48 GB VRAM | 43.2 GB | $6500 | $135 |
Rule of thumb: choose a GPU if you want speed on models up to about 30B. Choose a high-memory Mac if you want to load 70B-class models without a multi-GPU build. See the Quickstart grid