Every GPU and what it can do.
Memory bandwidth sets each card's ceiling; the board shows who gets close.
Memory bandwidth sets each card's ceiling; the board shows who gets close.
Cards with results come first. The ceiling is theoretical: bandwidth divided by the 3.4 GB that Qwen3.8 Flash-Next Q4_K_M reads per token.
| GPU | Bandwidth | Rigs | Best generation | Typical share of ceiling | Ceiling, Qwen3.8 Flash-Next Q4_K_M |
|---|---|---|---|---|---|
| Apple M3 Ultra | 819 GB/s | 0 | No runs | 238tok/s max | |
| Apple M2 Ultra | 800 GB/s | 0 | No runs | 233tok/s max | |
| Apple M1 Ultra | 800 GB/s | 0 | No runs | 233tok/s max | |
| Apple M5 Max | 614 GB/s | 0 | No runs | 179tok/s max | |
| Apple M4 Max | 546 GB/s | 0 | No runs | 159tok/s max | |
| Apple M5 Max 32-core GPU | 460 GB/s | 0 | No runs | 134tok/s max | |
| Apple M4 Max 32-core GPU | 410 GB/s | 0 | No runs | 119tok/s max | |
| Apple M3 Max | 400 GB/s | 0 | No runs | 116tok/s max | |
| Apple M2 Max | 400 GB/s | 0 | No runs | 116tok/s max | |
| Apple M1 Max | 400 GB/s | 0 | No runs | 116tok/s max | |
| Apple M5 Pro | 307 GB/s | 0 | No runs | 89.4tok/s max | |
| Apple M3 Max 30-core GPU | 300 GB/s | 0 | No runs | 87.3tok/s max | |
| Apple M4 Pro | 273 GB/s | 0 | No runs | 79.5tok/s max | |
| Apple M5 | 153.6 GB/s | 0 | No runs | 44.7tok/s max | |
| Apple M4 | 120 GB/s | 0 | No runs | 34.9tok/s max |