NumPy性能异常：旧MacBook为何比高端新ThinkPad快数个数量级？

我开发了一个重度依赖NumPy和Pandas的Python工具，意外发现它在全新高端ThinkPad上的运行速度远慢于一台使用十年的MacBook。为定位问题，我运行了以下基准测试：

import numpy as np
import time

def timeit(name, func, repeat=3):
    times = []
    for _ in range(repeat):
        start = time.perf_counter()
        func()
        times.append(time.perf_counter() - start)
    print(f"{name:<30} Min: {min(times):.4f}s  Avg: {sum(times)/len(times):.4f}s")

print("\n=== NumPy Benchmark ===")
print(np.__version__)
np.random.seed(42)

N = 2000

# Create large matrices for multiplication
A = np.random.rand(N, N)
B = np.random.rand(N, N)

timeit("Matrix multiply A @ B", lambda: A @ B)
timeit("Matrix multiply B @ A", lambda: B @ A)

# Basic math on large array
X = np.random.rand(N * N)

timeit("Elementwise add", lambda: X + 1)
timeit("Elementwise sin", lambda: np.sin(X))
timeit("Sort", lambda: np.sort(X))

# Optional: Linear algebra (solve Ax = b)
b = np.random.rand(N)

timeit("Solve Ax = b", lambda: np.linalg.solve(A, b[:N]))

MacBook测试结果 #

=== NumPy Benchmark ===

1.23.1

Matrix multiply A @ B          Min: 0.1090s  Avg: 0.1163s
Matrix multiply B @ A          Min: 0.1074s  Avg: 0.1152s
Elementwise add                Min: 0.0124s  Avg: 0.0161s
Elementwise sin                Min: 0.0375s  Avg: 0.0410s
Sort                           Min: 0.3826s  Avg: 0.3861s
Solve Ax = b                   Min: 0.1072s  Avg: 0.1127s

Windows ThinkPad测试结果 #

=== NumPy Benchmark ===
1.23.1
Matrix multiply A @ B          Min: 0.0897s  Avg: 0.1020s
Matrix multiply B @ A          Min: 0.0850s  Avg: 0.0923s
Elementwise add                Min: 0.0088s  Avg: 0.0090s
Elementwise sin                Min: 0.0260s  Avg: 0.0262s
Sort                           Min: 0.4298s  Avg: 0.4385s
Solve Ax = b                   Min: 4.1486s  Avg: 4.2807s

两台设备均运行Python 3.9及NumPy 1.23.1版本，其中线性方程组求解的性能差异最为显著。

两台设备的NumPy配置 #

MacBook配置 #

np.show_config()
openblas64__info:
    libraries = ['openblas64_', 'openblas64_']
    library_dirs = ['/usr/local/lib']
    language = c
    define_macros = [('HAVE_CBLAS', None), ('BLAS_SYMBOL_SUFFIX', '64_'), ('HAVE_BLAS_ILP64', None)]
    runtime_library_dirs = ['/usr/local/lib']
blas_ilp64_opt_info:
    libraries = ['openblas64_', 'openblas64_']
    library_dirs = ['/usr/local/lib']
    language = c
    define_macros = [('HAVE_CBLAS', None), ('BLAS_SYMBOL_SUFFIX', '64_'), ('HAVE_BLAS_ILP64', None)]
    runtime_library_dirs = ['/usr/local/lib']
openblas64__lapack_info:
    libraries = ['openblas64_', 'openblas64_']
    library_dirs = ['/usr/local/lib']
    language = c
    define_macros = [('HAVE_CBLAS', None), ('BLAS_SYMBOL_SUFFIX', '64_'), ('HAVE_BLAS_ILP64', None), ('HAVE_LAPACKE', None)]
    runtime_library_dirs = ['/usr/local/lib']
lapack_ilp64_opt_info:
    libraries = ['openblas64_', 'openblas64_']
    library_dirs = ['/usr/local/lib']
    language = c
    define_macros = [('HAVE_CBLAS', None), ('BLAS_SYMBOL_SUFFIX', '64_'), ('HAVE_BLAS_ILP64', None), ('HAVE_LAPACKE', None)]
    runtime_library_dirs = ['/usr/local/lib']
Supported SIMD extensions in this NumPy install:
    baseline = SSE,SSE2,SSE3
    found = SSSE3,SSE41,POPCNT,SSE42,AVX,F16C,FMA3,AVX2
    not found = AVX512F,AVX512CD,AVX512_KNL,AVX512_SKX,AVX512_CLX,AVX512_CNL,AVX512_ICL

ThinkPad配置 #

openblas64__info:

    library_dirs = ['D:\\a\\numpy\\numpy\\build\\openblas64__info']

    libraries = ['openblas64__info']

    language = f77

    define_macros = [('HAVE_CBLAS', None), ('BLAS_SYMBOL_SUFFIX', '64_'), ('HAVE_BLAS_ILP64', None)]

blas_ilp64_opt_info:

    library_dirs = ['D:\\a\\numpy\\numpy\\build\\openblas64__info']

    libraries = ['openblas64__info']

    language = f77

    define_macros = [('HAVE_CBLAS', None), ('BLAS_SYMBOL_SUFFIX', '64_'), ('HAVE_BLAS_ILP64', None)]

openblas64__lapack_info:

    library_dirs = ['D:\\a\\numpy\\numpy\\build\\openblas64__lapack_info']

    libraries = ['openblas64__lapack_info']

    language = f77

    define_macros = [('HAVE_CBLAS', None), ('BLAS_SYMBOL_SUFFIX', '64_'), ('HAVE_BLAS_ILP64', None), ('HAVE_LAPACKE', None)]

lapack_ilp64_opt_info:

    library_dirs = ['D:\\a\\numpy\\numpy\\build\\openblas64__lapack_info']

    libraries = ['openblas64__lapack_info']

    language = f77

    define_macros = [('HAVE_CBLAS', None), ('BLAS_SYMBOL_SUFFIX', '64_'), ('HAVE_BLAS_ILP64', None), ('HAVE_LAPACKE', None)]

Supported SIMD extensions in this NumPy install:

    baseline = SSE,SSE2,SSE3

    found = SSSE3,SSE41,POPCNT,SSE42,AVX,F16C,FMA3,AVX2

    not found = AVX512F,AVX512CD,AVX512_KNL,AVX512_SKX,AVX512_CLX,AVX512_CNL,AVX512_ICL

性能差异的核心原因 #

1. OpenBLAS的编译实现与优化程度不同

   • MacBook上的NumPy使用C语言实现的OpenBLAS，链接的是系统级预编译库，这类库针对Apple CPU架构做了深度优化，尤其是线性代数运算路径。
   • ThinkPad上的NumPy使用Fortran 77实现的OpenBLAS，且库路径指向numpy构建时的临时目录，说明这是未针对Intel/AMD CPU做定制优化的通用编译版本，缺少多线程、缓存调度等关键性能优化。

2. LAPACK求解器的效率差距
   np.linalg.solve依赖LAPACK库，MacBook上的LAPACK是集成在优化后OpenBLAS中的C语言实现，而ThinkPad上的是未优化的Fortran 77版本，无法充分利用CPU支持的SIMD指令集，导致线性方程组求解这类计算密集型任务性能暴跌。

3. 库的安装与链接方式差异
   MacBook的OpenBLAS是独立安装的系统级优化库，而ThinkPad上的库是numpy源码构建时默认生成的临时版本，没有经过完整的性能调优流程。

解决建议 #

• 在ThinkPad上通过conda install numpy重新安装NumPy，conda会提供针对Windows平台预编译的优化版OpenBLAS。
• 手动安装Intel MKL库，并配置NumPy链接到MKL，替代默认的OpenBLAS。

内容的提问来源于stack exchange，提问作者rhz