Oh i see, thanks that fits my case then!

Oh yeap, gotcha. They seem fine. The GPU for instance on the Cloud is around 60C.

I see, so this shouldn't be causing the issue you reckon.

Thanks for the input. I just did sudo dmesg --follow and then run my model and I don't see any errors. It just informs it loaded the UVM driver...

The memory is reasonable unless, of course, I match it close to the limit (e.g., batch size).

And what are the "temps"?

Thank you for your input, it makes sense. Nonetheless, I have adequate RAM and I just checked the IO speed (using sysbench) and actually are pretty much the same with the VM's being a bit faster.

> multithreaded
>
>sysbench
>
> to compare CPU and RAM

Thanks a lot for your input! I checked the CPU %steal it seems optimal ranging from 0.0 to 0.1st. Then, I don't think it's a shard since in the NVIDIA I have the full 80Gb memory at my disposal (unless they do some trickeries). I did a series of `sysbench tests and I found out that the VM's CPU is slightly worse for single-thread performance, but what is more astounding is the RAM speed. For 1 or 8 threads the write is 0.8x slower and the read is 1.5x slower. The Ram speed drop seems to reflect the iteration per second speed drop when I train the model. I guess this might be the fault.

Yes indeed, I am not doing anything in parallel. I use them separately and I wanted to compare their internal design as you said.

Thanks for your advice. FYI, I use the A100 for larger models and/or longer inputs/outputs that don't fit to my 3080.

I haven't looked into that. I would guess it wouldn't matter in my case, but I might be wrong.

I see, thanks! In that case, I might be asking the vendor more questions.