Decoding Google’s Cloud Gaming Announcement of Stadia

stadia.png

Google’s announcement of Stadia collided head on with scepticism from competitors and market participants. Microsoft’s prior announcements have been somewhat more vague, and did not have the direct and simple impact of Google’s Stadia debut at GDC. The announcement is a clear sign that disruptive forces are at work in game distribution. This is causing all players to examine their strategies and look toward forward looking solutions.

NGCodec is working on the biggest challenges in game streaming - from conventional streaming to AR/VR/XR - employing our ultra low latency encoding technologies on the latest, most capable codecs such as H.265, VP9 and AV1. Building the cloud gaming systems of the future will require improvements that lower latency, increase scalability and lower costs. For now, let's look at Google’s announcement and see what we might learn about Cloud Gaming’s near future.

Hardware Platform: Today, the system is in prototype mode, running on Intel-based servers and discrete AMD GPUs. Upcoming AMD semi-custom 7nm GPUs – not APUs – will make it possible to virtualize instances for more flexible performance and resolution scaling. This design future-proofs these servers – achieving more performance is a simple matter of allocating more CPU cores and potentially 2 GPUs to a single user instance. Future, faster hardware will allow fewer cores and back to a single GPU. AMD has been working on GPU virtualization for Stadia, and combined with an advanced video encoder their solution has to potential to hit multiple performance and resolution targets.

Video Encoder: The driver for a semi-custom GPU is likely the video encoder. Today the system is using AMD’s H.264 hardware encoder on prototype servers encoding to 1080p60 at approximately 25Mbps which is very high. I believe Google’s VP9 and AV1 hardware encoding (BigE and BigSea) technology is being incorporated into the semi-custom Stadia GPU to achieve 4K60, and designed properly, to allow multiple GPUs to achieve 8K and VR gaming. I likewise believe it is possible that AMD will keep this IP for their general market GPUs which will help VP9 and AV1 proliferate. VP9 and AV1 are  royalty free codec designed and championed by Google and AOM. We estimate around 10Mbps for 1080p60. The NGCodec VP9 RealityCodec has been deployed by Twitch and typically delivers good VQ at 4.5MBbps for 1080p60 with a 1 second latency. For sub frame latency we would need around 7Mbps for 1080p60.

Scaling: From a business perspective, this design permits multiple qualities of service, but the single GPU remains the big block. My guess is that the target Stadia Server will support 4 virtualized users with 4 GPUs and 2 CPUs which would still be possible in a 1U server chassis target heat/power. But I could be wrong – lots of design options exist from vertical high density stacks to higher density 2U and 4U designs. In a way it doesn’t matter – once Google solves the virtualization problem and GPU allocation, these options open.

Latency: Will be high. No getting around it. I estimate 150ms motion to photon with the cloud-edge loop, but we are developing detailed latency budget over the coming days. The challenge will be consumer WiFi. The latency of 802.11ac varies widely especially  when there is interference. My guess is that Google will encourage use of Ethernet. You really want 5G or wired fibre optics Internet for this – and relatively computationally expensive video encode/decode, exacerbated by a hodgepodge of decoders not latency optimized. TV screens have notoriously high latency HDMI ports. Over time Google and the industry can address this, but for now, the digital world is not optimized for low latency cloud gaming.

Console: One debate that has likely brewed inside Google is whether to embrace the Hybrid model of Microsoft. While unlikely, Google could encase the Stadia hardware in plastic and sell a console with identical to cloud performance and zero developer changes. Wait – Google is launching a cloud platform – why would it launch a game console? The truth is, lots of users have unreliable connectivity and may fear what happens if they do not maintain an active subscription. A console hedges the bet, and they are already doing all work to build the hardware and software platform anyway. But they don’t want to take the wind out of their cloud announcement so look for the console to come later, if at all.

NGCodec currently demonstrates game streaming with virtual reality at 90FPS at the native resolution of the HTC Vive using our RealityCodec solution. Reach out to simon.solotko@ngcodec.com if you would like to learn more.