Hey, just working on optimising my demo using Unreal Insights. Revisiting the question I had in my mind on Animation performance, “new” Motion Matching vs. “old” State Machines & Blendspaces. Sharing my findings below for anyone else that was interested in that question.
Context
I profiled a traditional state machine animation blueprint (our game, BGP) against UE 5.7’s GASP sample using motion matching via Unreal Insights single-frame timer analysis. Both traces had full stat named events enabled.
The Key Numbers
| State Machine (BGP) | Motion Matching (GASP) | |
|---|---|---|
| Worker thread eval per character | ~0.43 ms | 0.50 ms |
| Locomotion selection cost per character | 0.008 ms | ~0.25 ms |
| Game thread anim cost per character | ~0.07 ms | ~0.14 ms |
| Debug cost (strippable) | 0.003 ms/char | 0.016 ms |
What This Means
- Motion matching locomotion selection is ~31x more expensive than state machine evaluation (0.25 ms vs 0.008 ms per character)
- BUT overall worker thread cost per character is surprisingly comparable (~0.43 vs 0.50 ms) because each system spends its budget differently:
- BGP fills its worker budget with PoseDriver (0.21 ms/char) and ControlRig (0.08 ms/char) - procedural corrections that would remain regardless of locomotion approach
- GASP fills its worker budget with motion matching pipeline (0.25 ms/char) - pose search, trajectory, choosers, BlendStack
- Both test scenarios were GPU-bound, meaning worker thread animation had headroom and did not bottleneck the frame
- If BGP adopted motion matching while keeping PoseDriver/ControlRig, net per-character worker cost would increase by ~0.24 ms (from ~0.43 to ~0.67 ms)
- At 19 ticking characters, that’s ~4.6 ms additional aggregate worker thread time (parallel, not blocking game thread), reducible via URO
The Biggest Single Cost in Either System
PoseDriver in BGP at 3.94 ms total (0.21 ms/char, 210 calls across 19 characters). This is not related to state machines or motion matching - it’s a procedural correction system that would exist in either architecture. It is the single most expensive animation timer in our entire trace.
Test Methodology
Environment
- Engine: Unreal Engine 5.7
- Profiling tool: Unreal Insights, single-frame timer list export with stat named events enabled
- Hardware: Same machine for both captures
BGP Tutorial Area (State Machine)
- Rich game world with environment, NPCs, lighting, particles
- 19 animation-ticking skeletal meshes (leader components running ABPs)
- 426 total skeletal mesh component ticks (modular character system - each character has multiple body part meshes following a leader pose)
- Animation blueprint uses traditional state machine with blendspaces, plus IK (LegIK, ControlRig), PoseDriver nodes, layered blending, montage slots
- Game thread frame: ~9.3 ms
- Parallel animation evaluation enabled (28 worker tasks)
GASP Motion Matching Sample
- Simple scene, single controllable character
- 1 animation-ticking skeletal mesh
- Animation blueprint uses motion matching (PoseSearch), Chooser Tables, BlendStack, orientation warping, steering, foot placement
- No PoseDriver, no ControlRig (for animation)
- Game thread frame: ~5.0 ms (3.2 ms idle/GPU-wait)
- Parallel animation evaluation enabled (1 worker task)
Normalization Approach
BGP costs are divided by character count (19 for most timers) to produce per-character estimates. GASP costs are taken directly (1 character). Worker thread aggregate times represent total CPU time across all workers, not wall-clock game thread impact.
Game Reference Link - The Freeblades on Steam