reduction of taichi is slower than numpy

[Frenchfries135]

import taichi as ti
import numpy as np
import time
ti.init(arch=ti.cuda)#,debug=True)
n=1<<18
a=np.random.randn(n).astype(np.float32)
b=ti.field(ti.f32,n)
b.from_numpy(a)
@ti.kernel
def sum1()->ti.f32:
    result=0.0
    for i in b:
        result+=b[i]
    return result
#compile
sum1()
t1=time.time()
for _ in range(1000):
    a.sum()
t2=time.time()
for _ in range(1000):
    sum1()
ti.sync()
t3=time.time()
print('time numpy:',t2-t1,'time ti atomic_add:',t3-t2)
input('enter to continue')

result in RTX3060 and R7 5800H is：

time numpy: 0.07301616668701172 time ti atomic_add: 0.3090794086456299