Performance
All three features run continuously while enabled, so the question that matters to a developer is what they take from the machine during play. This page gives measured cost and latency on four reference configurations.
The models are small classifiers distributed in ONNX format and executed through ONNX Runtime, with hardware acceleration supplied by Windows ML. Placement is decided automatically per machine, and the player configures nothing. Windows ML execution providers require Windows 11 24H2 or newer; below that the platform falls back to what it can reach without them.
Reference configurations
Four machines were measured, chosen to differ from each other.
| A | B | C | D | |
|---|---|---|---|---|
| Form factor | Laptop | Desktop | Desktop | Laptop |
| CPU | Intel Core i9-14900HX | Intel Core Ultra 5 250K Plus | Intel Core Ultra 7 270K Plus | AMD Ryzen AI 9 HX 370 |
| Discrete GPU | NVIDIA RTX 4090 Laptop | AMD Radeon RX 9070 XT | AMD Radeon RX 7900 XT | None |
| Integrated GPU | Intel UHD Graphics | Intel Graphics | Intel Graphics | AMD Radeon 890M |
| NPU | None | Intel, via OpenVINO | Intel, via OpenVINO | AMD Ryzen AI, via VitisAI |
Between them they cover NVIDIA and AMD discrete cards, Intel and AMD processors, an Intel NPU and an AMD Ryzen AI NPU, and one machine with no NPU at all.
On every machine that has one, the discrete GPU is left idle by the platform. No RTX 4090, Radeon RX 9070 XT or 7900 XT is asked to run any of this work.
This is deliberate. Pinned measurements show the CPU is often the fastest device for models this small, and the comparison is not close. Production selects accelerator hardware regardless, because keeping the CPU available to the game during play is worth more than reducing an inference time that already has generous headroom.
During a game session, inference runs on the NPU or the integrated GPU, both of which sit idle while the game is on the discrete card. Even in the runs below that forced everything onto the discrete GPU, the average cost stayed under 2% of that GPU for every feature.
Session cost by execution path
Steady-state process usage sampled every 500 ms, across every execution path available on each machine. Each cell gives CPU % average / GPU % average, where the GPU is the one that path runs on. The final row of each table is the production selection, which is the configuration a player actually gets; it is measured as a separate run with nothing forced, so it does not always match a single row above it.
Machine key. A: Intel Core i9-14900HX, RTX 4090 Laptop, Intel UHD Graphics, no NPU. B: Intel Core Ultra 5 250K Plus, Radeon RX 9070 XT, Intel Graphics, Intel NPU. C: Intel Core Ultra 7 270K Plus, Radeon RX 7900 XT, Intel Graphics, Intel NPU. D: AMD Ryzen AI 9 HX 370, Radeon 890M integrated, Ryzen AI NPU, no discrete GPU.
Impact Engine
| Execution path | A | B | C | D |
|---|---|---|---|---|
| CPU | 0.2 / 0.0 | 0.2 / 0.0 | 0.1 / 0.0 | 0.1 / 0.0 |
| CPU, OpenVINO | 1.1 / 0.0 | 1.4 / 0.0 | 1.1 / 0.0 | N/A |
| Discrete GPU, DirectML | 0.1 / 1.8 | 0.1 / 0.3 | 0.0 / 0.4 | N/A |
| Discrete GPU, TensorRT RTX | 0.1 / 2.8 | N/A | N/A | N/A |
| Discrete GPU, MIGraphX | N/A | N/A | N/A | N/A |
| Integrated GPU, DirectML | 0.1 / 3.9 | 0.1 / 4.9 | 0.1 / 2.1 | 0.1 / 1.1 |
| Integrated GPU, OpenVINO | 0.2 / 3.1 | 0.3 / 22.7 | 0.1 / 21.8 | N/A |
| Integrated GPU, MIGraphX | N/A | N/A | N/A | N/A |
| NPU, OpenVINO | N/A | N/A | 1.1 / 0.0 | N/A |
| NPU, VitisAI | N/A | N/A | N/A | N/A |
| Production selection | 0.2 / 3.2 | 0.3 / 22.5 | 0.1 / 21.9 | 0.1 / 0.0 |
Screen Sync
| Execution path | A | B | C | D |
|---|---|---|---|---|
| CPU | 1.8 / 0.2 | 2.2 / 0.0 | 1.6 / 0.0 | 3.1 / 0.0 |
| CPU, OpenVINO | 1.6 / 0.1 | 2.2 / 0.0 | 1.3 / 0.0 | N/A |
| Discrete GPU, DirectML | 2.0 / 0.2 | 2.1 / 0.0 | 1.4 / 0.0 | N/A |
| Discrete GPU, TensorRT RTX | 1.7 / 0.2 | N/A | N/A | N/A |
| Discrete GPU, MIGraphX | N/A | 2.0 / 0.1 | 1.5 / 0.0 | N/A |
| Integrated GPU, DirectML | 1.7 / 0.2 | 2.3 / 0.1 | 1.4 / 0.1 | 3.1 / 0.0 |
| Integrated GPU, OpenVINO | 1.8 / 0.2 | 1.9 / 0.7 | 1.7 / 0.7 | N/A |
| Integrated GPU, MIGraphX | N/A | N/A | N/A | 3.2 / 0.0 |
| NPU, OpenVINO | N/A | 1.6 / 0.1 | 1.5 / 0.1 | N/A |
| NPU, VitisAI | N/A | N/A | N/A | 2.9 / 0.1 |
| Production selection | 1.6 / 0.2 | 2.1 / 0.1 | 1.4 / 0.1 | 3.4 / 0.1 |
Mood Analyzer
| Execution path | A | B | C | D |
|---|---|---|---|---|
| CPU | 0.9 / 0.0 | 0.8 / 0.0 | 0.8 / 0.0 | 1.4 / 0.0 |
| CPU, OpenVINO | 1.5 / 0.0 | 1.0 / 0.0 | 1.0 / 0.0 | N/A |
| Discrete GPU, DirectML | 1.2 / 0.5 | 0.7 / 0.1 | 0.7 / 0.1 | N/A |
| Discrete GPU, TensorRT RTX | 1.1 / 1.0 | N/A | N/A | N/A |
| Discrete GPU, MIGraphX | N/A | N/A | N/A | N/A |
| Integrated GPU, DirectML | 1.3 / 1.1 | 0.7 / 1.4 | 0.7 / 0.6 | 1.4 / 0.3 |
| Integrated GPU, OpenVINO | 1.0 / 0.9 | 1.5 / 5.8 | 0.9 / 5.5 | N/A |
| Integrated GPU, MIGraphX | N/A | N/A | N/A | N/A |
| NPU, OpenVINO | N/A | N/A | 1.0 / 0.1 | N/A |
| NPU, VitisAI | N/A | N/A | N/A | N/A |
| Production selection | 0.9 / 0.8 | 2.4 / 5.7 | 2.0 / 6.0 | 1.6 / 0.0 |
N/A means the path produced no measurement on that machine, either because the hardware or provider is not present or because the run did not complete. Machine D has no OpenVINO rows at all, since OpenVINO targets Intel devices and that machine is an AMD platform.
Figures use the float variant. The int8 variant differs by only a few tenths of a percent at session level, so the picture is the same either way.
CPU cost stays low on every machine and every path, between 0.1% and 3.4% on average in the production configuration. Screen Sync carries the highest of the three, since it is the only feature doing image work.
GPU load depends almost entirely on which integrated GPU is present and where the work was placed, and on every machine with a discrete card that integrated GPU is otherwise idle during play. Impact Engine uses 22.5% of the integrated GPU on machine B and effectively none on machine D, where it runs on the CPU instead. Read the GPU column as the cost to a component the game is not using, and as specific to a configuration.
Since Screen Sync and Mood Analyzer are mutually exclusive, a running system carries at most one of those two plus Impact Engine. The figures should not be added across all three.
Inference latency
Per-inference cost of each model on its production device, warm cache, average.
| Model | Drives | A | B | C | D |
|---|---|---|---|---|---|
| Audio event classifier | Impact Engine, Mood Analyzer | 10.21 ms | 7.16 ms | 4.76 ms | 4.22 ms |
| Mood classifier | Mood Analyzer | 3.46 ms | 3.95 ms | 3.45 ms | 1.06 ms |
| Screen Sync vision, first stage | Screen Sync | 6.08 ms | 4.95 ms | 4.67 ms | 1.54 ms |
| Screen Sync vision, second stage | Screen Sync | 5.94 ms | 5.02 ms | 4.91 ms | 0.89 ms |
Compare these against the cadences in Signal characteristics. Every model completes well inside the interval between its analyses, with the audio classifier using a small fraction of its budget and the vision stages finishing in a few milliseconds. The smaller models run fastest on machine D.
The analysis cadence is therefore set by design rather than by what the hardware can keep up with.
All of these measurements come from four configurations, taken in a benchmark host process with no game running alongside. Machine A's latency figures come from a newer build than B, C and D. Read them as the shape of the cost rather than as a guarantee for a given title on arbitrary hardware.
Next
- Overview, feature summary and runtime model