Audio Performance

UE5 has three audio threads:

• Game thread — spawns ActiveSounds, fires notifies.
• Audio thread — bookkeeping; ActiveSound update loop.
• Audio render thread — DSP, mixing, output.

The audio thread is not the audio render thread — community confusion is constant; Epic staff repeatedly clarify on the forums. Blocking in a sound notify hits the game thread (because notifies fire there); blocking in DSP hits the render thread.

MaxChannels on most platforms defaults to 32, capped 1–64 internally (FAudioPlatformSettings). MaxChannels cannot exceed MaxSources — engine logs "Can't increase MaxChannels past MaxSources" if violated.

Devs trying to bump max voices to 64 in UE5 hit this and get a non-obvious clamp warning. Audit Project Settings → Audio → Max Channels per platform.

MetaSounds are sample-accurate DSP graphs — the modern path. SoundCue is the UE4 legacy. New projects: MetaSound; legacy projects: gradual migration.

MetaSound Pages (5.6+) let you ship a single MetaSound asset with multiple pages — quality-tiered graph variants — switched at runtime via au.MetaSound.Pages.SetTarget or per-platform device profile. The same MetaSound can ship a high-quality version for desktop and a stripped-down version for Switch without two assets.

Constructor pins in MetaSound graphs are read-only at runtime but cheaper than dynamic inputs — move inputs that don't need to vary onto constructor pins as a documented MetaSound optimization.

UE 5.7 added MetaSound Operator Cache memory reductions and Vorbis decode-only build option for memory-constrained targets.

Sound Concurrency is the asset that decides "how many of THIS sound type can play simultaneously." Settings:

• Max Voice Count — hard cap.
• Resolution Rule — what happens when a new sound exceeds the limit: StopOldest, StopFarthest, StopFarthestThenOldest, PreventNew.
• Generation Volume Scaling — older sounds quieter to make new ones audible.
• Voice Stealing Fade — gentle stop instead of hard cut.

A common shipping bug: a designer removes a Sound Concurrency asset, and active voice count climbs from 38 to 61 silently. Invisible in stat unit, obvious in audio-thread profiling.

When a sound goes inaudible (out of range, too quiet to hear), it can either be killed or virtualized — tracked in time without rendering audio:

• PlayWhenSilent — the sound continues in time even when silent. Useful for music sync, looping ambience. CPU cost continues.
• Seek Restart (5.7+) — new mode that restarts the sound at the right point when it becomes audible again, without paying CPU during silence.

The trap: PlayWhenSilent on a long looping MetaSound is a steady CPU cost that doesn't show up in any HUD. Audit which sounds need PlayWhenSilent and which can use Seek Restart.

Attenuation falloff curves, occlusion traces, reverb send methods — all per-voice CPU you can't see in stat unit. au.SoundDistanceOptimizationLength sets the max duration a one-shot must be in order to be a candidate for distance-based culling. Short stingers may bypass culling and pile up.

Convolution reverb — block size is the actual cost lever; defaults are not free, IR length is hidden CPU. Higher block size = lower CPU + higher latency; lower block size = opposite. Hardware acceleration is optional per platform.

Source effects run per-voice. CPU multiplies by active voice count. Apply only when a single voice needs unique processing (per-character voice modulation).

Submix effects run on a shared mixed bus. Pay once for any number of voices feeding the submix. Apply for global effects (master EQ, master reverb, mastering compressor).

The trap: replicating a submix effect chain as source effects on every voice. Multiplies the CPU cost by the voice count. Audit on a per-voice basis whether the effect is genuinely source-specific.

If you ship Wwise, the cost shape changes. Audiokinetic's blog explicitly calls out:

• AkComponent per-frame work — every registered AkComponent does work each frame. Spawning a new one per event instead of reusing existing components is the most common Wwise+UE perf bug.
• SoundBank streaming — oversized banks blow load times. Audiokinetic recommends category-based banks loaded on demand.
• Wwise 2025.1 Command Buffer API reduces per-frame integration overhead.
• The Wwise integration was rewritten in 2022.1 — older tutorials (4.27 era) describe a different code path.

Tools:

• stat audio — audio thread + mixer summary.
• stat soundwaves / stat soundcues / stat soundmixes / stat soundsources — per-asset-type counters.
• au.Debug.Sounds 1 / au.Debug.SoundCues 1 — per-frame text overlays.
• au.3dVisualize.Enabled 1 — in-world audio gizmos.
• Insights audio trace + 5.6 MetaSound asset scopes for build/render memory.

Common audio perf killers:

• Designer removes Sound Concurrency asset; voice count silently climbs.
• PlayWhenSilent on a long looping MetaSound — steady invisible CPU.
• Source effects replicating what should be a submix chain.
• Convolution reverb with too-short block size + long IR.
• AkComponent spawn-per-event instead of reuse.
• 4K UI textures... wait, that's a different tutorial.

PerfGuard can baseline active voice count + audio thread time per scenario, gate CI on regressions like "active voices climbed from 38 to 61" or "audio render thread went from 1.4 ms to 3.2 ms after a feature merge." Surface MetaSound Operator Cache memory deltas via 5.6+ Insights asset scopes.

• Memory & VRAM Optimization — audio memory budgeting.
• Performance Gotchas — cross-cutting issues.