Demystifying Geforce Now: KubeVirt & Openshift
After working on my self hosted cloud gaming : which you can read here
Gauransh Mathur
I wanted to get it working for me when playing outside on mobile data, but it’s not as easy as it looks.
Difference between wifi and mobile data
Now wifi is hard wired from the router. Latency is determined by the wifi speed and your connection to the router, so if you just put an ethernet cable into your device you just fixed everything.
But mobile data is different, you are connecting to multiple mobile towers based on movement and based on your position. With 5G your speed increases but you won’t notice that latency is still high.
Yes there is a difference
When you are watching a video or TikTok you are using UdP. This allows packet loss as well as unstable connections because each data packet is not so important, keeping the connection is.
Latency comes because of multiple reasons, you can’t wire a mobile data so distance, signal strength, congestion all affect it causing you to not be able to have a stable connection.
CGNAT
This is your biggest problem maker. ISP’s need a way to give you a public ip without them having to spend too much. Wifi is simple as it’s a singular point where you have control. Mobile data has too many changes.
CGNAT stands for carrier grade network address translation. This forces your traffic to flow in the most cost effective way to be able for the ISP to reduce costs forcing many people to have the same ip. You can still bypass this using VPN or Relays or things like Tailscale. But you still won’t have a perfect connection and still some latency overhead.
Geforce Now
Nvidia has done a lot of cool things with Geforce now, recently upgrading their own maintained servers to blackwell architecture.
Nvidia also provides partnerships with alliance partners for other regions, which don't carry the most powerful gpu's, but if you are streaming from handhelds or from a phone / tablet, it won't really matter because you will be able to max out the gpu.
GPU's
The GPU's themselves are not consumer cards. These are same generation enterprise cards that have the ability to do Virtualisation. There are 3rd party non supported hacks around to do virtualisation on a consumer card but it is unstable.
You will also notice, compared to the consumer card alternative, the gpu in the cloud is given more VRAM this allows the card to push further than normal.
Numa
Non unified memory access, this is nvidia's main way to be able to reduce latency in the games itself. NUMA makes sure the CPU core is attached to the correct GPU PCIe lane. CPU cores have dedicated lanes that in a consumer system doesn't matter because you aren't putting multiple CPU's and GPU's together. But in a server making sure you bind the CPU to the correct lane allows you to reduce latency by 50ms maybe even more sometimes.
KubeVirt & OpenShift
KubeVirt is an extension of kubernetes that allowes you to run VM's and manage them like pods. It requires some kernel level extensions and works basically via CRD's. This is the main method that Nvidia uses to be able to run their VM's. They can start any VM in Linux or Windows in kiosk mode so you only see 1 game and they can manage how many VM's are running, shutdown, created or even deleted.
Openshift is their main software to run kubernetes, this is an enterprise grade k8s that is run by RedHat. You can use others but RedHat's version has a lot more customisability and its specifically to manage the whole cluster
I have even managed to try the GFN Ultimate from Japan servers to test the 4080 and 5080 and even with high latency its still very playable.
But I will be switching back to using GFN in singapore as I get a discount as a starhub user and the usage being only when I'm out, allows me to still keep my self-hosted cloud gaming for when I'm at home
