How Taobao’s “Future City” Guarantees 200ms Cloud‑Rendered 3D Experiences

Background

Amid the metaverse boom, Taobao launched a 3D immersive shopping scene called Future City during the 2022 Double‑11 event, leveraging powerful cloud GPUs to deliver cinematic‑grade graphics with an end‑to‑end latency of about 200 ms, allowing users to explore, shop, and socialize without downloading large packages.

Technical Architecture

Future City replaces the traditional push‑stream / pull‑stream model with a 1‑to‑1 streaming session for each user. When a user opens the entry page, a scheduler assigns an idle cloud Windows instance. The instance runs a Unreal Engine client integrated with the ARTC SDK, converting the user’s avatar and actions into rendered frames that are streamed back to the client.

Content creation: 3D avatars, animations, buildings, goods, environments.

User side: Taobao/Diantao app with ARTC SDK for joystick and gesture control.

Cloud rendering host: Windows VM with Unreal Engine and ARTC SDK handling rendering, encoding, and streaming.

Cloud rendering scheduler: Manages lifecycle of rendering instances, recycles resources on idle or abnormal exit.

Business services: User lifecycle management and gameplay logic.

Multi‑user interaction service: Broadcasts user actions via ACCS for real‑time interaction.

Streaming center: Combines playback control, ARTC, and GRTN to deliver the video stream and upstream command channel.

Quality‑Assurance Challenges

Early development revealed several problems: missing release‑process checks, lack of stable test environments, configuration drift between pre‑release and production machines, no restart strategy for cloud hosts, and the absence of a one‑click rollback mechanism.

Redefined Release & Test Process

The team built a multi‑environment pipeline (development, test, gray‑verification, production) and introduced online test‑group nodes for gray verification. The revised flow includes:

Pre‑release environment: functional, multi‑user interaction, and performance tests must pass and be product‑approved before any release.

Upload cloud game: the uploaded package must match the approved pre‑release version.

Online test group: simulate multi‑user load, re‑validate product acceptance, and ensure worst‑case device performance.

Online deployment: staged rollout in at least two groups without disrupting connected users.

Online rollback: support graceful restart after users disconnect or forced restart for critical issues, with batch rollback capability.

Performance Standards

Key objective metrics include:

Control latency < 200 ms.

First frame displayed within 2 seconds.

Stable 30 FPS playback.

Rendering quality criteria cover avatar completeness, realistic clothing, correct body posture, accurate 3D goods appearance, and physically plausible scene lighting and collisions.

Multi‑User Interaction Testing

Robots simulate multiple avatars using the AOI (Area of Interest) algorithm to limit visible users, then broadcast aggregated interaction data via ACCS. Performance is measured with Unreal Engine commands and memory‑leak checks during repeated clothing swaps.

Cloud Host Baselines

Across hardware configurations, the team established thresholds:

draws < 2000.

CPU usage < 15 %.

GPU memory per instance < 3.5 GB.

Monitoring & Incident Management

Post‑release, a real‑time dashboard tracks crash rates, latency, and first‑frame times. Automated DingTalk bots broadcast anomalies to developers, enabling rapid response. Crash rate dropped from 1.14 % to 0.23 % and first‑frame latency doubled after optimizations.

Common Rendering Defects

Typical issues observed during testing include avatar clipping, missing clothing textures, stuck character movement, jagged or blurry edges, lighting flicker, and color shifts caused by incorrect material rendering.

Future Outlook

The quality‑assurance framework will continue evolving as cloud rendering technology matures, with ongoing refinement of standards, tools, and performance targets to support richer immersive commerce experiences.