实验过的环境如下，结论都成立：

但是据同事在Win10的Linux子系统下验证，据说 将numpy转为pytorch的tensor后反而比前者更慢 ，怀疑是子系统实现产生的问题。

下面是验证流程。

import time
import numpy as np
import torch
print(torch.__version__)
1.4.0
def check_time(func, run_times=10):
    t = time.time()
    for i in range(run_times):
        func()
    print('avg time = %s sec' % ((time.time()-t)/run_times))
shape = (5000,5000)
a = np.ones(shape, dtype=np.float)
b = np.ones(shape, dtype=np.float)
k = np.ones((shape[0],1), dtype=np.float)
# - simple numpy ndarray plus
def test_np_1():
    c = a+b
    return c
check_time(test_np_1)
avg time = 0.21692438125610353 sec
# - simple numpy ndarray and broadcast
def test_np_2():
    c = a+b+k
    return c
check_time(test_np_2)
avg time = 0.45278918743133545 sec
# - use pytorch tensor
def test_torch_1():
    ta = torch.from_numpy(a)
    tb = torch.from_numpy(b)
    tc = ta+tb
    c = tc.numpy()
    return c
check_time(test_torch_1)
avg time = 0.11778402328491211 sec
# - use pytorch tensor and broadcast
def test_torch_2():
    ta = torch.from_numpy(a)
    tb = torch.from_numpy(b)
    tk = torch.from_numpy(k)
    tc = ta+tb+tk
    c = tc.numpy()
    return c
check_time(test_torch_2)
avg time = 0.2651021957397461 sec
# - check gpu
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
ga = torch.from_numpy(a).float().to(device)
gb = torch.from_numpy(b).float().to(device)
gk = torch.from_numpy(k).float().to(device)
cuda:0
# - try tensor on gpu
def test_torch_cuda_1():
    ca = torch.from_numpy(a).float().to(device)
    cb = torch.from_numpy(b).float().to(device)
    cc = ca+cb
    c = cc.cpu().numpy()
    return c
check_time(test_torch_cuda_1)
avg time = 0.44039239883422854 sec
# - try tensor on gpu and broadcast
def test_torch_cuda_2():
    ca = torch.from_numpy(a).float().to(device)
    cb = torch.from_numpy(b).float().to(device)
    ck = torch.from_numpy(k).float().to(device)
    cc = ca+cb+ck
    c = cc.cpu().numpy()
    return c
check_time(test_torch_cuda_2)
avg time = 0.4477779150009155 sec
# - try tensor on gpu and broadcast, preload in gpu before call, and not copy to cpu after
def test_torch_cuda_3():
    cc = ga+gb+gk
    return cc
check_time(test_torch_cuda_3)
avg time = 0.0004986286163330078 sec