Benchmarks
How much work can Tide's ECS do in a tick? These small CPU workloads give a baseline for comparing ECS performance.
This measures ECS updates, including deferred structural changes where used. It excludes setup, world hashing, snapshots, networking and rendering. It does not measure the cost of a complete Tide multiplayer tick or predict a game's frame rate.
Milliseconds per ECS tick · lower is faster · bars start at zero
Bars show medians. Lines and smaller values show the range of process medians.
4 processes per implementation and workload · 200 warmup ticks · 400 measured ticks per process
Versions and implementations
- Tide fe5e96be1 — Generated C storage and scheduled systems; moving Frozen tags. Clang 23.1.2, -O2, baseline x86-64.
- Bevy 0.20.0 — Table storage and scheduled systems; Rust 1.95.0, opt-level 2, baseline x86-64.
- Flecs 4.1.6 — Cached archetype systems and pipeline; Clang 23.1.2, -O2, baseline x86-64.
- EnTT 4.0.0 — Views and an owning filtered group with explicit dispatch, without an integrated scheduler; Clang 23.1.2, -O2, baseline x86-64.
| Workload | Threads | Tide | Bevy | Flecs | EnTT |
|---|---|---|---|---|---|
| 1M particles | 1 | 0.446480.43545–0.57870 | 1.155631.14835–1.26580 | 1.667101.66455–1.73415 | 1.883971.84365–1.91285 |
| 1M wide components, all active | 1 | 1.291371.22825–1.34430 | 2.408372.38005–2.60395 | 2.259552.24280–2.28730 | 2.853602.82575–2.88020 |
| 1M wide components, 50% frozen | 1 | 0.445220.42905–0.46295 | 1.143581.10590–1.20100 | 1.103471.09495–1.12315 | 1.465601.43635–1.46895 |
| 1M wide components, 95% frozen | 1 | 0.029250.02890–0.02930 | 0.101200.10000–0.10690 | 0.109630.10910–0.11205 | 0.131950.13005–0.13720 |
| 1M wide components, 99% frozen | 1 | 0.005750.00570–0.00580 | 0.020220.02010–0.02060 | 0.021870.02165–0.02250 | 0.029030.02670–0.02970 |
| 2k bullets, churn | 1 | 0.346950.33555–0.38910 | 0.520380.46550–0.68610 | 0.568750.45420–0.73950 | 0.262670.25980–0.27045 |
Tide has the lowest sampled median for particles, dense wide components and the three frozen densities. EnTT has the lowest churn median. Bevy and Flecs churn ranges are wide; these observations describe this run, not a universal ranking.
The 96 accepted processes use four cyclic rounds: each implementation occupies every launch position once per workload. One complete attempt that overlapped a build was excluded; its raw evidence and the accepted replacement are both retained. Attempts were not selected by their timing.
Before timing, 75 validation cases matched across the four native implementations, an independent C reference and Tide WASI; the harness passed 26 checks. The raw report records source and binary hashes, correctness evidence and the attempt ledger. Bevy, Flecs and EnTT WebAssembly were not measured or validated in this report.
Reading the results
Lower is faster. Each cell is the median of separate process medians, in milliseconds per tick. The smaller numbers underneath give the minimum and maximum of those process medians. These ranges show run-to-run variation; they are not confidence intervals. Close or overlapping results need more measurements before calling a winner.
Every workload stays visible, including cases where another implementation is faster. There is no overall score: moving particles, filtering wide components and changing entity structure exercise different costs. Reports from different runs remain separate.
What the workloads do
| Workload | Work measured |
|---|---|
| 1M particles | Update position and velocity for one million particles. |
| 1M wide components | Update one million entities with wider component data. |
| 50%, 95% or 99% frozen | Update only the remaining active entities. Frozen membership is fixed during the run; this does not measure changing filter membership. |
| 2k bullets, churn | Repeated spawning and destruction alongside movement. Deferred changes are applied inside the timed tick. |
The implementations use the same initial values and update arithmetic. The runner checks final state digests across implementations before accepting a result. That checks these workloads; it does not establish cross-platform determinism for the other ECS libraries.
What is being compared
Tide generates specialized C storage and systems for each game. Bevy uses its Table component storage and schedule. The broader serial comparison also uses Flecs cached archetype systems and EnTT views and an owning filtered group with explicit dispatch. EnTT does not supply an integrated scheduler in this comparison, so these paths do not provide identical engine services.
Each report names its exact implementations, versions and thread count. Compiler settings preserve the update's floating-point operation order: no fast math or fused multiply-add contraction. Tide keeps its determinism requirements; the harness does not disable them to improve a result.
These are selected synthetic workloads on one machine, maintained by the Tide project. Cache layout, entity count, query shape, structural changes, scheduling and hardware can change the outcome. Use them as a starting point for a workload that resembles your game.