When getting images through jpeglivesource, relation between nvdec and GPU memory.

MIP
, ,
1

Hello there

I’m developing deep learning based software.

By its nature, the software use much GPU memory.

At first, nvdec worked sometimes, I didn’t know why.

Recently, I found out that if usable memory is lower than 20% of the total memory that the GPU has nvdec stops.

this limitation, is it from nvidia driver or MIP SDK?

is there any method that I can work around?

thanks.

2

The NVidia decoder implementation by Milestone has some load balancer logic and it will stop using the Nvidia GPU if the memory utilization is above 80 percent. This is a hardcoded limit that you cannot change, the limit is based on experience while developing and testing at Milestone. I guess this fit perfectly with your observations.

3

ok, then…

my problem is that.

if I use 2GB or 4GB of memory GPU, it doesn’t matter. or up to 8GB. it is ok. I can accept that.

But the problem is, if I use Tesla T4 (16GB), Tesla V00(32GB), Quadro RTX5000(16GB), RTX6000(24GB) or RTX8000(48GB) it’s huge problem.

​

guess 20% of them.

if it is RTX8000 it’s almost 10GB.

I think you should reconsider your policy for these high-end GPUs.

4

Thank you for your feedback. I consulted an expert in Milestone Development and they will take it consideration but we cannot say when (or even if) this will be developed as there can also be larger memory fragmentation even if there is a lot of memory available.