Ascend NPU Benchmark
Atlas 800I A2
FLUX.1-dev
Ulysses: Standard Ulysses Attention, Async Ulysses: Ulysses Attenton with Async QKV Projection, NPU Attn Backend: NPU Attention Backend, UAA: Ulysses Anything Attenton
| 800I A2x1 |
800I A2x2 w/ Ulysses |
w/ Async Ulysses |
w/ Async Ulysses + NPU Attn Backend |
| FLUX.1, 16.13s |
🎉11.45s |
10.47s |
🎉10.34s |
 |
 |
 |
 |
Qwen-Image-Edit
| 800I A2x2 w/ UAA + TEP |
800I A2x4 w/ UAA + TEP |
w/ UAA + TEP + Async Ulysses |
w/ UAA + TEP + Async Ulysses + NPU Attn Backend |
| Qwen-Image-Edit, 134.76s |
🎉67.44s |
64.82s |
🎉64.43s |
 |
 |
 |
 |
Z-Image-Turbo w/ NPU Attn Backend
| 800I A2x1 |
800I A2x2 w/ Ulysses |
w/ Async Ulysses |
| Z-Image-Turbo, 3.39s |
🎉2.49s |
🎉2.38s |
 |
 |
 |