CANN/asnumpy 基准测试
Benchmarks
【免费下载链接】asnumpy哈尔滨工业大学计算学部苏统华、王甜甜老师团队联合华为CANN团队开发的华为昇腾NPU原生Numpy仓库项目地址: https://gitcode.com/cann/asnumpy
Back to README | Reproduce:
examples/03_multiply.py
This document contains the full performance benchmark comparing AsNumpy (NPU) against NumPy (CPU) on themultiply()operation.
Test Environment
| Item | AsNumpy (NPU) | NumPy (CPU) |
|---|---|---|
| Processor | Ascend 910B NPU | Server CPU (AArch64) on the same machine |
| NPU Runtime | CANN 8.2.RC1.alpha003 | — |
| Python | Python 3.9+ | |
| Library version | AsNumpy 0.2.0 | NumPy 1.26+ |
| Data type | float32 | |
| Operation | multiply() — element-wise multiplication | |
| Timer | time.perf_counter() (high-resolution) | |
Controlled Variables
- Both sides useidentical input data: arrays are generated by NumPy and transferred to NPU via
from_numpy()before timing starts. - Data transfer time is excluded: only the
multiply()computation is timed. - Results are single-run wall-clock times (no warmup, no averaging).
Results
| Shape | AsNumpy (NPU) | NumPy (CPU) | Speedup |
|---|---|---|---|
| (500, 500) | 1.9355 s | 0.1708 s | 0.09× |
| (1000, 1000) | 0.0692 s | 0.7029 s | 10.16× |
| (2000, 2000) | 0.1033 s | 3.8387 s | 37.17× |
| (3000, 3000) | 0.1115 s | 14.3567 s | 128.70× |
Key observation:For small tensors (500×500), NPU launch overhead dominates and CPU is faster. As tensor size grows, NPU's massive parallelism takes over — reaching128.70× speedupat 3000×3000.
Reproducing the Results
Run the benchmark script from the project root:
python examples/03_multiply.pyThe script tests all four shapes with 50 iterations each, reports average and minimum times, and verifies numerical correctness against NumPy (relative diff < 1e-4).
【免费下载链接】asnumpy哈尔滨工业大学计算学部苏统华、王甜甜老师团队联合华为CANN团队开发的华为昇腾NPU原生Numpy仓库项目地址: https://gitcode.com/cann/asnumpy
创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考