Did some quick tests after building llama.cpp with ROCm 6.4.2 and latest Vulkan for my 6900 XT
gemma4 E2B Q4_K
| ubatch | ROCm pp512 | Vulkan pp512 | ROCm tg128 | Vulkan tg128 |
|---|---|---|---|---|
| 32 | 1536.60 | 1423.49 | 151.92 | 174.59 |
| 64 | 1590.65 | 1930.60 | 151.41 | 173.76 |
| 128 | 2651.11 | 2998.42 | 151.53 | 173.71 |
| 256 | 3653.19 | 3233.44 | 151.45 | 173.45 |
| 512 | 3807.60 | 3950.71 | 151.47 | 173.67 |
| 1024 | 3806.77 | 3948.27 | 151.49 | 173.35 |
qwen35 4B Q8_0
| ubatch | ROCm pp512 | Vulkan pp512 | ROCm tg128 | Vulkan tg128 |
|---|---|---|---|---|
| 32 | 1368.32 | 706.18 | 77.57 | 88.58 |
| 64 | 1841.68 | 1323.46 | 77.65 | 88.57 |
| 128 | 2577.95 | 1672.51 | 77.97 | 88.46 |
| 256 | 2984.38 | 2244.62 | 77.72 | 88.50 |
| 512 | 3023.75 | 2390.09 | 77.81 | 88.57 |
| 1024 | 3019.70 | 2386.97 | 77.60 | 88.53 |
[link] [comments]