Split rendering

Split rendering is a way of producing the images for a virtual reality (VR) or augmented reality (AR) device in which the graphics workload is divided between the device and a more powerful computer reached over a network, such as a PC, a phone, an edge server or a cloud data center. The remote machine renders the demanding 3D scene and streams the result as compressed video, while the headset or glasses keep the time-critical work, above all tracking and a final correction of the image to the user's latest head pose. 3GPP's study of extended reality over 5G describes the basic form as one in which the server "rasterizes the XR viewport and does XR pre-rendering", while the device is "able to do latest pose correction, for example by asynchronuous time-warping".

The closely related term remote rendering is older and broader. Shu Shi and Cheng-Hsin Hsu, in a 2015 survey in ACM Computing Surveys, define it as "rendering 3D graphics on a computing device and displaying the results on another computing device connected through a network." In XR, the word "split" stresses that part of the rendering stays on the device. The technique lets small, light or battery-powered devices show graphics they could not render on their own, and it underlies wireless PC VR streaming such as Air Link, cloud services such as NVIDIA CloudXR and the retired Azure Remote Rendering, and AR glasses built on the Qualcomm Snapdragon AR2 Gen 1, which stream from a phone or other host device.

Reviewed 15 October 2026. Checked every claim against 3GPP TR 26.928 V16.0.0, TS 26.565 V18.3.0, the Qualcomm 2018 deck, Mark's 1999 dissertation, Hou et al. 2017, Crossref records for the cited papers, Microsoft Learn pages and the cited Road to VR, UploadVR, MIXED, VentureBeat, Purdue and NVIDIA articles. About review dates.

Definition and terminology

Shi and Hsu note that remote rendering was first developed for sharing computing resources remotely, and that interest grew with cloud computing and mobile devices. In XR the term covers a range of designs that differ in how much work stays on the device.

3GPP Technical Report 26.928, Extended Reality (XR) in 5G, separates several cases:

• Viewport rendering in the network: the server renders the whole viewport from the tracking and sensor data it receives, and the device only decodes the video and displays it "without using the viewport information". The report says this allows simple clients but has significant compression and transport challenges, because every step from tracking to display has to fit within a 20 ms motion-to-photon limit.
• Raster-based split rendering: the server renders the viewport, and the device corrects the image to the latest pose, for example with asynchronous timewarp (ATW). The XR graphics workload is "split into rendering workload on a powerful XR server (in the cloud or the edge) and pose correction (such as ATW) on the XR device".
• Generalized XR split rendering: the server pre-renders the 3D scene into "a simpler format" that may carry extra metadata, and the device "does the final rendering based on local correction on the actual pose". The report says this may give more flexibility in bitrates, latency and processing than single-buffer split rendering, and that buffers other than 2D frame buffers could be split over the network.

The report concludes that split compute and rendering is a new delivery category for 5G, alongside the conventional conversational, interactive, streaming and download categories.

How it works

The 3GPP report gives the basic procedure for raster-based split rendering:

1. The XR device connects and sends static information about itself, such as its supported decoders and its viewport.
2. The server sets up its encoders and formats to match.
3. In a loop, the device collects its pose (or a predicted pose) and sends it to the server.
4. The server uses that pose to pre-render the viewport, encodes it with 2D video encoders and sends the compressed video back together with the pose it was rendered for.
5. The device decodes the video and uses the pose that came with the frame and its own current pose to correct the image locally, for example with ATW.

The final step is what separates split rendering from plain video streaming. In a 2018 presentation on what it called boundless XR, Qualcomm showed "on-device adjustment to latest pose" on the headset, with partial rendering offloaded to an edge cloud over 5G. The deck states that asynchronous time warp "reduces Motion to Photon (MTP) latency by using on-device processing based on the latest available pose", and that MTP below 20 ms, which generally avoids discomfort, "has to be processed on the device." Ericsson described the same division in 2020: graphics rendering on the edge cloud augments head, controller, hand and motion tracking that stay on the headset because they are latency-sensitive.

The device can also combine remote and local content. Microsoft's documentation for Azure Remote Rendering says that "at the end of a frame, your locally rendered content is automatically combined with the remote image", so an application could draw its own interface on the device while the cloud rendered a large model.

The idea of correcting an already rendered image for a new viewpoint is older than current XR hardware. William R. Mark's 1999 dissertation at the University of North Carolina at Chapel Hill, supervised by Gary Bishop, rendered every Nth frame normally and produced the frames in between with a 3D image warp that is "approximately correct for a new viewpoint and view direction". Mark wrote that the technique "can also compensate for rendering-system latency and network latency", and he built a real-time system that displayed rendered images at a remote location with low latency. The conference paper "Post-rendering 3D warping" by Mark, Leonard McMillan and Bishop appeared at the 1997 Symposium on Interactive 3D Graphics.

Latency and bandwidth

Version 16.0.0 of TR 26.928 gives these working figures for raster-based split rendering:

Parameter	Value given in TR 26.928
Motion-to-photon latency	At most 20 ms, met by processing inside the device when ATW is used
Pose-to-render-to-photon (end-to-end) latency	50 ms (clause 6.2.5.5); described as "typically 50-60ms" in clause 6.2.5.1
Frame buffer format	2k by 2k per eye, potentially higher
Frame rate	At least 60 fps, potentially up to 90 fps
Downlink bitrate	In the order of 50 Mbps per eye buffer with H.264/AVC in experiments; 50 to 100 Mbps given as a valid target
Uplink	Mainly pose information, several hundred kbit/s
Suggested network QoS	New 5QIs addressing latency in the range of 10 to 20 ms with guaranteed bitrates of 50 to 100 Mbps

The report adds that split rendering "requires the use of edge computing", since the pose-to-render-to-photon time is expected to be below 50 ms. A 2017 paper by Xueshi Hou, Yao Lu and Sujit Dey of the University of California, San Diego, which proposed light VR glasses that connect wirelessly to edge or cloud computers performing the rendering, analyzed what it called the "challenging bitrate and latency requirements" of that approach.

History

Date	Development
1997	Mark, McMillan and Bishop present post-rendering 3D warping at the Symposium on Interactive 3D Graphics.
2015	Shi and Hsu publish their survey of interactive remote rendering systems.
July 2017	Hou, Lu and Dey present "Wireless VR/AR with Edge/Cloud Computing" at ICCCN.
October 2017	Researchers from Purdue University and Tsinghua University present Furion, a split renderer for phone VR, at ACM MobiCom 2017.
September 2018	Qualcomm presents split rendering over 5G as "boundless photorealistic mobile XR".
25 February 2019	Microsoft announces Azure Remote Rendering with HoloLens 2.
22 October 2019	NVIDIA announces CloudXR for streaming cloud-rendered AR and VR over 5G.
25 February 2020	Qualcomm and ZeroLight show an end-to-end enterprise system with edge cloud rendering over on-premises 5G.
2020	3GPP TR 26.928 V16.0.0 (Release 16) identifies split compute and rendering as a new delivery category; ETSI publishes it in November 2020.
14 April 2021	Oculus announces Air Link, wireless PC VR streaming for Oculus Quest 2.
16 November 2022	Qualcomm announces the Snapdragon AR2 Gen 1 for glasses that stream from a host device.
June 2023	Azure Remote Rendering enters public preview on Quest 2 and Quest Pro.
July 2025	ETSI publishes 3GPP TS 26.565 V18.3.0, the Release 18 Split Rendering Media Service Enabler.
30 September 2025	Microsoft retires Azure Remote Rendering.
17 March 2026	NVIDIA announces native integration of CloudXR 6.0 in visionOS for Apple Vision Pro.

Applications in VR and AR

PC to headset streaming

Streaming from a local PC to a standalone headset is a consumer form of remote rendering. Oculus announced Air Link in April 2021 to provide the PC VR streaming of its wired Oculus Link over Wi-Fi on Quest 2. It launched as an experimental feature, and Oculus advised using a 5 GHz router in the same room with the PC connected to the router by Ethernet. When Valve released Steam Link for Meta Quest in December 2023, MIXED described it as the fourth PC VR streaming solution for the headsets, alongside Virtual Desktop, Air Link and the open-source ALVR.

For AR, Microsoft's Holographic Remoting runs an app on a PC instead of on a HoloLens. Inputs from the HoloLens, namely gaze, gesture, voice and spatial mapping, are sent to the PC, which renders the content and sends the rendered frames back to the headset. Microsoft lists two uses: previewing and debugging apps during development, and letting an app use a PC's resources when it has high-resolution assets or models. Remote apps for HoloLens 2 use version 2.x of the remoting package and are not compatible with HoloLens (1st gen).

Phone or host to glasses

Qualcomm's Snapdragon AR2 Gen 1, announced on 16 November 2022, was designed for glasses that do not run applications themselves. UploadVR reported that "AR2 isn't designed to power standalone devices": a smartphone or compute puck runs the AR apps and streams the output wirelessly to the glasses over a FastConnect 7800 Wi-Fi 7 link, for which Qualcomm claimed a stream latency of less than 2 milliseconds.

Edge and cloud

Azure Remote Rendering was Microsoft's cloud service for rendering complex 3D models and streaming them to HoloLens 2 or Windows desktop PCs from Unity or native C++ apps. Microsoft illustrated the need with a side-by-side engine model: a high-fidelity version with more than 18 million triangles and a reduced version with around 200,000, the kind of simplification required on untethered devices with less rendering power. The service rendered the model on cloud GPUs, encoded the image as a video stream and, for models too large for one GPU, spread the work across several GPUs and merged the result into one image. Microsoft first announced it on 25 February 2019 with HoloLens 2, added a public preview for Quest 2 and Quest Pro in June 2023, and retired the service on 30 September 2025.

NVIDIA announced CloudXR in October 2019 as a set of tools for businesses to stream cloud-rendered AR and VR content over 5G, supporting SteamVR and OpenVR content. NVIDIA claimed "no detectable latency difference" compared with local rendering; Road to VR's Ben Lang was skeptical of that claim and noted that the physical distance to the data center matters for latency as much as the network. In March 2026 NVIDIA said CloudXR 6.0 was natively integrated in visionOS, and that dynamic foveated streaming, enabled by CloudXR 6.0 and visionOS 26.4, keeps the highest visual fidelity at the user's focus points while saving performance and bandwidth.

Network operators and chipmakers have demonstrated edge split rendering for enterprise use. In February 2020 Qualcomm and ZeroLight showed a system in which an XR2-based headset handled tracking and some low-power, latency-sensitive tasks, while a ZeroLight rendering server on the customer's premises, using NVIDIA hardware and CloudXR, streamed a photorealistic car configurator over a millimeter-wave 5G network operating entirely within the enterprise's premises. VentureBeat reported that the XR2 design included Ericsson-validated 5G and that Qualcomm was working with Goertek on industrial designs for the hardware. Ericsson described a similar setup combining the Snapdragon XR2 5G reference design, NVIDIA RTX GPUs running CloudXR at the edge, and a 5G user plane placed "as close to the user as possible".

Research

Furion, presented by Zeqi Lai, Y. Charlie Hu and colleagues at ACM MobiCom 2017, is a split renderer for phone VR. It renders the background on a PC or server and the foreground on the phone. According to Purdue University, a Pixel XL rendering on its own managed only 111 ms per frame, and full server rendering streamed over Wi-Fi took around 200 ms per frame; with the split, frame rendering time fell to 14 ms on the Pixel XL, and the paper's abstract reported under 14 ms latency at 60 FPS, the phone's display refresh rate. A journal version appeared in IEEE Transactions on Mobile Computing in 2020.

Hou, Lu and Dey's 2017 paper started from the observation that headsets of the time were heavy and that rendering was done locally on a tethered phone, PC or console, and examined how light glasses could instead connect wirelessly to edge or cloud computers that perform the rendering.

Standardization

3GPP's TR 26.928 proposed, as a short-term step, developing "a framework and basic functionalities for raster-based Split Rendering for Online Gaming", followed in the mid-term by a more general split and cloud rendering framework. A later Release 18 specification, TS 26.565, defines a Split Rendering Media Service Enabler covering functions on the user's device and on the media application server, plus an API that lets app developers start and manage split rendering sessions. Its architecture pairs a Split-Rendering Client on the device with a Split-Rendering Server; the client may reach the XR runtime through OpenXR or WebXR, and the client sends pose metadata and receives rendered media over a WebRTC session. The specification names immersive gaming and immersive communication with avatars as typical use cases, and says it targets XR applications but may also be used for 2D displays.

Limitations

The sources above identify several costs. 3GPP notes that network-rendered viewports put heavy demands on compression and transport, and Hou, Lu and Dey describe the bitrate and latency requirements of wireless VR as challenging. Road to VR noted that latency depends on the physical distance to the data center as much as on the network. When Oculus introduced Air Link, it said the wired Oculus Link would still give the best quality and most reliable experience.

See also

• Edge computing
• NVIDIA CloudXR
• Air Link
• Timewarp
• Reprojection
• Motion-to-photon latency
• 5G
• Foveated rendering
• Rendering