Peer access test

[turboderp]

import torch, time

tensor = torch.zeros((4*1024, 4*1024), dtype = torch.int, device = "cuda:0")

print(f"Torch says peer access is: " + "enabled" if torch.cuda.can_device_access_peer else "disabled")

for j in range(10):

    a = time.time()
    for i in range(100):
        tensor2 = tensor.to("cuda:1")
        torch.cuda.synchronize()
    b = time.time()
    print(f"Peer copy: {b - a:.4f} seconds")

    a = time.time()
    for i in range(100):
        tensor2 = tensor.to("cpu").to("cuda:1")
        torch.cuda.synchronize()
    b = time.time()
    print(f"CPU copy: {b - a:.4f} seconds")

[jmoney7823956789378]

A4000 x2:
Torch says peer access is: enabled
Peer copy: 1.0322 seconds
CPU copy: 3.9984 seconds
MI60 x2
Torch says peer access is: enabled
Peer copy: 0.5041 seconds
CPU copy: 3.4114 seconds

[EyeDeck]

Here's a result with some mismatched cards in case it's interesting.

3090 (PCI-E 4.0 x16) + 3080 10GB (PCI-E 4.0 x4):

Torch says peer access is: enabled
Peer copy: 2.3494 seconds
CPU copy: 3.4054 seconds
Peer copy: 2.2270 seconds
CPU copy: 3.4658 seconds
Peer copy: 2.2222 seconds
CPU copy: 3.4598 seconds
Peer copy: 2.2233 seconds
CPU copy: 3.4513 seconds
Peer copy: 2.2234 seconds
CPU copy: 3.4769 seconds
Peer copy: 2.2250 seconds
CPU copy: 3.3917 seconds
Peer copy: 2.2246 seconds
CPU copy: 3.4589 seconds
Peer copy: 2.2250 seconds
CPU copy: 3.4451 seconds
Peer copy: 2.2227 seconds
CPU copy: 3.4558 seconds
Peer copy: 2.2226 seconds
CPU copy: 3.4478 seconds