for (int i = 0; i<gpus; ++i)
        for (int j = 0; j < gpus; ++j)
        {
            if (i != j)
            {
                cudaDeviceCanAccessPeer(&access, i, j);
                if (access == 1)
                    cout << "T: " << i << " -> " << j << "n";
                else
                    cout << "N: " << i << " -> " << j << "n";
            }
        }