cpu_gpu_copy.py

import torch
import numpy as np

import time

elapsed = 0


for _ in range(10):
  a1 = np.random.random((8192, 8192)).astype(np.float32)
  b1 = torch.tensor(a1)
  st = time.time()
  c1 = b1.cuda()
  elapsed += time.time() - st
  
bandwidth =  a1.nbytes / (elapsed / 10) / 1E9
print(f"Total time: {elapsed}")
print(f"GBPS: {bandwidth}")