Yi Huan (Perfect World) launched April 29 — Unreal Seoul tech talk

Liu Jenyang opened with the problem that looks simple from the player seat: walk across a district, step into a phone booth, and appear somewhere else almost instantly. For the team that behavior is anything but trivial. Their core rule was “trigger early, handle late”: spawn a virtual camera and streaming sources at the target location before the player arrives, then fetch terrain and data in the background so visible loading never interrupts play. Teleport and zone changes themselves are gated in three explicit stages. Gate one verifies safe loading of the destination data, gate two checks that the play environment can be recovered, and gate three returns control to the player only after dynamic objects such as NPCs and vehicles and the local world state are fully resolved. Liu said that sequence prevents incomplete world states from ever being exposed to players.

Memory management for overlapping zones — the so-called dual-residency window where both origin and destination live in memory — required strict budgeting. Rather than promote every object to a heavy full actor, the team used lightweight placement representations to minimize both memory footprint and activation cost during transitions. Large parts of the city are filled with actors that represent everyday life, so the studio balanced which entities stayed lightweight and which became full actors only when necessary. That selective activation lowered peak memory use and helped avoid out-of-memory conditions during handoffs between areas.

A headline technical win came from crowd and entity optimization. Starting on Unreal Engine 5.5 and carrying that work into 5.6, Hota Studio adopted the Mass framework to manage thousands of pedestrians and vehicles as low-cost data units. The tricky moment is when many NPCs simultaneously request heavy computation in the same frame. To avoid hitching, the team deliberately delays spawning full actors even when entities enter view; streaming and collision conditions must be fully satisfied before actor creation. They also take a single view-state snapshot to evaluate entities during rapid camera rotation, then apply triple buffering to prevent data races. Around that core they built a three-layer runner budget: Update Rate to throttle logic frequency, Pending Range to limit how far they evaluate candidates, and Submission Caps to cap how many actors get spawned per frame. On Android devices they simulated 300 agents and reported that applying submission-cap control sharply reduced frame drops and kept performance close to 60fps.

Feng Tianli followed with the studio’s scalability policy for bridging PC and mobile hardware differences. He emphasized that the mobile pipeline is not just the PC sliders turned down; it is a separate, purpose-built stack that accounts for prolonged thermal load, throttling and bandwidth limits. To support Android fragmentation the team layered device profiles from broad platform families down to specific model mappings, and they prioritize per-model rules over generic GPU rules. They added a verification-driven trust system: even when a device is tagged with an Epic or Cinematic profile, runtime checks across GPU vendor, rendering path and CPU classification must pass before Lumen or hardware ray tracing turn on.

Device coverage uses a two-tier model. A basic fallback layer provides a safety net that should work across all devices, while a higher-coverage tier targets 90–95% of intended devices via more than 390,000 detailed profiles. Feng also noted a curated whitelist for some hardware quirks: although Vulkan is Android’s default API, 135 devices are on a separate whitelist where OpenGL ES or feature-disable modes are more stable. Those exceptions let the studio avoid enabling expensive features on devices that can’t handle them reliably.

Directional light and shadow handling — a heavyweight cost in any open city — got a series of practical optimizations. The day-night cycle is split into 12 segments and the sun’s position is held fixed within each segment. That stability allows the team to avoid redrawing distant shadow maps every frame and safely reuse cached shadow data across multiple frames. On PC they tackled cache invalidation caused by giant instanced objects: instead of invalidating the whole shadow cache when a large bounding box crosses the view, they convert the camera frustum into a 2D footprint and update only the actually visible tiles. Liu reported that change cut unnecessary full-shadow updates by about 90% and reduced GPU shadow cost by 20–30%.

Mobile shadowing uses a hybrid pipeline: a responsive CSM for near-range detail, and a layered shadow map (LSM) for distant shadows that locks to camera position rather than view direction. Because LSM doesn’t follow camera rotation, it avoids re-rendering when the player turns, and region moves are split into N×N tiles that get refreshed over several frames. The system also reuses shadow pixels from the previous region and tracks the mesh IDs that contributed to the cache so only newly appearing objects need full re-rendering. Feng closed his talk with capture footage from an iPhone 16 Pro and the line that framed their engineering goal: maximize cache reuse and keep all updates inside the frame budget.