| Chip | Memory bandwidth | Max memory | NPU | Largest model it holds |
|---|---|---|---|---|
| Apple M5 Max | 460-614 GB/s | 128 GB | n/a | gpt-oss 120B |
| Apple M5 Pro | 307 GB/s | 64 GB | n/a | Llama 3.3 70B |
| NVIDIA GB10 Grace Blackwell | 273 GB/s | 128 GB | n/a | gpt-oss 120B |
| AMD Ryzen AI Max+ 395 | 256 GB/s | 128 GB | 50 TOPS | gpt-oss 120B |
| Qualcomm Snapdragon X2 Elite Extreme | 228 GB/s | 128 GB | 80 TOPS | gpt-oss 120B |
| Apple M6 | 153-170 GB/s | 32 GB | n/a | Qwen3.6 35B-A3B |
| Apple M5 | 153 GB/s | 32 GB | n/a | Qwen3.6 35B-A3B |
| AMD Ryzen AI 9 HX 370 | 89.6 GB/s | 256 GB | 50 TOPS | gpt-oss 120B |
"Largest model it holds" assumes the chip's biggest memory option, a long (8,000-token) conversation and 4-bit models from our calculator's catalog. Memory bandwidth is the maker's theoretical peak; see our bandwidth study for how it translates into speed.