Could you please provide reference results?

[rsdmse]

Thank you for providing such a great tool. I'm wondering if you could also provide the results for various GPU architectures, because we're not sure if our GPUs are configured optimally and sometimes I find it difficult to reproduce the benchmarks provided by NVIDIA.

I ran the device_to_device_memcpy_read_ce test on 6 A6000 devices and got the following result:

nvbandwidth Version: v0.8
Built from Git version: v0.8

CUDA Runtime Version: 13000
CUDA Driver Version: 13000
Driver Version: 580.95.05

...

memcpy CE GPU(row) -> GPU(column) bandwidth (GB/s)
           0         1         2         3         4         5
 0       N/A     26.34     26.35     21.95     22.64     21.80
 1     26.34       N/A     26.34     21.86     22.62     22.63
 2     26.34     26.35       N/A     22.36     21.87     22.17
 3     21.82     21.82     22.26       N/A     26.34     26.35
 4     22.64     22.64     21.81     26.35       N/A     26.34
 5     21.86     22.67     22.01     26.34     26.35       N/A

SUM device_to_device_memcpy_read_ce 715.53

Does this look about right to you?

[kunaldeo]

Without NVLink these look good to me. Almost 80% of full PCIe Gen 4 speeds.