Hi, We want to build an Unreal Pixel Streaming web server.
We design the infrastructure so that at least 20 clients can connect and use it at the same time.
We want to build in a legacy environment, not in the cloud.
When we tested it with our own ‘NVIDIA RTX A4000 D6 16GB’, we were able to stream up to 4 instances.
What graphics card (GPU) should we use to run more than 20 instances?
Hi, Welcome to the NVIDIA Developer forums! This is the forum feedback category, unfortunately there is no one from support monitoring this section. I am moving this to the GPU Hardware category for better exposure.
Hello @lhh7135 !
Will this design be a many to many application in the sense that every client will have its own application instance running on the server? Or rather a multi-client setup where only one app runs on the server but gets streamed to many users? I think you target the first one, but I would like to make sure.
I ask because those are different scenarios. In the latter case it is only about encoding capabilities, while the first puts much more load on the GPU because it needs to render for every client.
The RTX A4000 is already a top tier Workstation GPU and the capabilities you need for pixel streaming many to many do not scale linearly. That means you might not be able to get 8 instances even with an RTX A 6000, which is roughly double the amount of processing power.
For your purposes you will need to look into multi-GPU setups or GPU servers like our DGX offerings.
In that context, should I try and connect you to our Enterprise services regarding these?
Thanks for your reply.
We plan to stream as one app. (1 to many)
We tested up to 4 instances on the A4000.
Referring to the test results, when configuring A5500 or A6000 x3 SLI, wouldn’t it be possible to stream 20 instances?
Is there any bottleneck of pixel streaming according to SLI configuration?
I assume with one app for many users you mean that the app is rendering an individual view for every user, correct?
As I stated before, you cannot expect this to scale linearly even when moving from A4000 to A6000. That means combining 3 GPUs does not necessarily guarantee 20 instances. It depends a lot on the actual workload and on how Unreal can parallelize across several GPUs.
The only expected overhead related to the SLI setup could be mem copies between GPUs since likely only one GPU will be encoding the outgoing video stream. And if you keep this at normal resolutions and refresh rates this should not be an issue.
Sadly I could not find internal reference data for such a scenario.
In summary, if Unreal supports Multi-GPU setups for pixel streaming and scales with the number of GPUs, then this is a good approach. And with 3xA6000 you should be able to if not quite reach 20 instances, at least get close to it.
Thanks again for the reply.
We want multiple users to access from one server and receive different results from each. However, according to your answer there was no reference to it and we understood that you should hope that Unreal’s pixel streaming is optimized for multi-GPU.
You are welcome.
This is correct, we don’t have internal reference setups for combining Workstation GPUs used for Unreal pixel streaming.
NVIDIA does of course have Game streaming technology, see our GeForce Now offerings. But these are cloud based and use dedicated Multi-Node/Multi-GPU servers.
Hi, please if you allow me to participate in the matter, I am looking for something very similar to the above mentioned, but different from it, I need the first scenario, (many to many applications), I also need to scale a hardware for 20 users in simultaneous, each with its separate instance, I saw that you have a new type of hardware specific to Android, with capacity for up to 160 games at the same time, (https://www.nvidia.com/en/data-center/rtx-server-gaming), what kind of hardware is available today for this scenario of 20 users or more for pixel streaming, can you help?