I have a code with heavy use of dataframes and lots of computation. I want to speed up the process time so I bought a Nvidia GPU and try to implement cudf library. But performance is awful. benchmark code below run very slow on gpu.

import cudf as cd
import pandas as pd
import numpy as np
import time
import cupy as cp

# Generating random data
np.random.seed(42)
size = 1000
iteration = 1000

data = {
    'type': np.random.choice(['X', 'Y'], size=size),
    'num1': np.random.randint(10, 1001, size=size),
    'num2': np.random.randint(1000, 100001, size=size)
}

# Creating DataFrame
df_cpu = pd.DataFrame(data)
df_gpu = cd.DataFrame(data)

for i in range (2):

    t = time.time()
    for i in range(iteration):
        temp = sum(df_cpu[df_cpu.type == 'X']['num1'].values)
    print(f'filtering in cpu in {time.time()- t}')

    t = time.time()
    for i in range(iteration):
        temp = sum(df_cpu.loc[df_cpu.type == 'X', 'num1'].values)
    print(f'filtering with loc cpu in {time.time()- t}')

    t = time.time()
    for i in range(iteration):
        temp = sum(df_gpu[df_gpu.type == 'X']['num1'].values)
    print(f'filtering in gpu in {time.time()- t}')

    t = time.time()
    for i in range(iteration):
        temp = sum(df_gpu.loc[df_gpu.type == 'X', 'num1'].values)
    print(f'filtering with loc in gpu in {time.time()- t}')

Benchmark results

nvdia-smi

Modules

What am I missing? I need to full utilize gpu and increase performance