Wowza Community

CUDA gone, and now?

Hi,

I’m using version 3.6 atm, but just ordered keys for 4.x.

I noticed that CUDA is gone for hw transcoding, which is very bad for me. The GPU I’m using doesn’t support NVENC.

How do I get the newest Wowza software, but still keep the CUDA support? I don’t have a CPU that can handle the load whatsoever.

I assume that this is the latest 4.0.4 Development build that I need? But where can I download it?

Secondly,

As all now needs to be based on NVENC or Intel QuickSync, what GPU should I replace my existing GPU with? I don’t have like $2000 for this, but need a very good GPU only for transcoding.

I have 6x 1080p live streams, so its really heavy for Wowza it seems.

Thanks

Kent

Hi Kent,

CUDA is still available in Wowza Streaming Engine 4.x however it has been deprecated and although there are no immediate plans, it may be completely removed at some stage. The reason is that the latest Nvidia drivers no longer support the version of CUDA hardware that Wowza was designed to use.

You should be able to continue using your CUDA hardware for the foreseeable future however, you will need to use a Nvidia driver that is earlier than version 340. If you use the Nvidia driver search, it will show the latest versions available that support your hardware. In most cases, this will be version 331.xx.

Roger.

Hi,

Thank you for your reply.

So CUDA is fully supported in 4.0.5 aswell? At least as long as I’m not updating my GPU drivers?

As for QuickSync it doesn’t seem to offload the CPU as much as CUDA. Not sure why, maybe because I have 6 streams?

NVENC is only supported on really highend GPU Cards. I don’t really have $2000 for a new GPU. It’s simply too expensive. K4000 is the cheapest that I could find.

Hi,

Thanks. I have just upgraded to 4.0.6 and waiting for cameras to get online. 6 ingoing RTSP streams with 1080p@30fps is alot more than the benchmarks show from wowza. It could be nice to have some updated stats on this :slight_smile:

Will it even be possible to run a full software decoding and encoding without hw acceleration? The reason I’m reencoding is that I have an overlay dynamic picture (scoreboard) that I need to apply.

I have considered using text instead of a small picture, but the stream will still need reencoding. So I don’t think it will help on performance which is a huge problem basically.

Thx

/kent

Also, either QuickSync or NVENC should provide similar or better performance as you currently have with CUDA.

With QuickSync, you are limited to a single processor but with NVENC, you can use multiple GPUs (if the motherboard supports it).

Roger.

Hi Kent,

Yes, QuickSync will show that it uses more CPU than CUDA because the encoder is built into the CPU whereas the CUDA encoder is completely separate.

Both perform about the same but where CUDA or NVENC has the advantage is that you can have multiple GPUs in the same server if the motherboard supports this.

A quick google search did locate K4000 GPUs for around the $1k mark which is similar to what the older M series GPUs used to go for and there is possibly still good resale value on your current GPU.

Another, possibly cheaper, option is to have a multi CPU server with lots of cores and memory and use the default (software) encoding method.

Roger.

Hi Kent,

decoding 6 x 1080p @ 30fps will put quite a bit of load on the CPU in any case. If QuickSync hardware available then that decoding load can be offloaded to the QuickSync GPU but it isn’t possible with CUDA or NVENC. If lots of CPU cores are available then it shouldn’t be an issue to decode and encode in software mode.

The type of overlay you use won’t really affect performance as each has to be converted to a YUV frame to be overlayed anyway. There isn’t any way to embed a graphic or text into a video without re-encoding it.

As I’ve already mentioned, there are no immediate plans to remove CUDA support completely but it may happen at some time. It has been listed as deprecated to let users know to look at the other options when sourcing hardware for transcoding.

Roger.