Inside the Adaptive Immersion Platform: On-Device AI That Reads Your Game

It is already live in Razer Chroma for Windows: players can simply select the immersive features from the Chroma dashboard and adjust them per device, app, and game.

The AI models analyze the media they digest, down sampled screen or audio frame, and run inference for the output they are trained for.

The end user can turn on and off these modes in the Chroma App today. Screen Synch and Mood Analyzer are mutually exclusive as they both control chroma; Impact Engine can run alongside them.

Alongside latency, data privacy is fundamental to maintain the trust of the gamers. No data leaves the PC of the user.

We also see in the future the rise of NPUs, which will become ubiquitous on the gamer's device. It was a no-brainer for us to focus where the action happen.

Doing so would require classifying what’s going on in the content and associate a haptic effect, this approach has latencies in the 100s of milliseconds, which are incompatible with gaming.

Impact Engine is a small on-device classifier listening to the system audio output continuously and identifies what kind of sound are present in the mix and their relative strength, driving audio haptics conversion based on this classification.

The intensity changes are shaped asymmetrically. Impacts rise instantly with no attack ramp. Tails decay gradually to avoid a choppy experience.

We tuned the impact engine to generate the best gaming experience with a wide range of genres, and as a background on which designed haptics effects can shine.

Music is treated the same way, which keeps a soundtrack from driving the haptics all game long. Where music is the point, in a rhythm game or a set-piece built around the score, there are two routes. A curated Audio-to-Haptics profile can apply gain on specific frequency intervals so the track itself carries the haptics, or the WYVRN SDK can author them explicitly.

We applied Razer’s 15 years of experience in Chroma RGB design to develop the best color model possible, complementing the experience.

Speech gets filtered out, so spoken content produces almost no reaction by design. And the result settles over a short window rather than reacting to every bar. Emotional state moves on a scale of seconds, so the lighting moves that way too.

We are quite proud of this one as it is where we started developing all our local inference models and it looks awesome.

Those rates explain how quickly the system reacts. For a developer, however, they raise the next practical question: if this analysis is running continuously while the player is in your game, how much of the machine does it use?

The first useful question is not simply how much compute the models need, but where that compute runs. In particular, a game developer wants to know whether these features compete with the discrete GPU that is already rendering the game.

Where the work goes instead depends on which silicon is available on the device. The work goes to the NPU where the machine has one, and to the integrated GPU or CPU where it does not.

Two things follow. Acceleration improves without a Razer release, since a better provider arriving through the platform is picked up automatically. And NPU coverage is not tied to one vendor, with Intel and AMD Ryzen AI both covered today.

Windows ML execution providers require Windows 11 24H2. Below that it is handled on the generic hardware without specific silicon acceleration.

The following measurements show steady-state process usage across all execution paths on each of the four machines, sampled every 500 ms. Each cell gives CPU % average, then iGPU % average.

N/A means the path produced no measurement on that machine, either because it is not present or because silicon execution provider was unable to compile the model for the silicon.

The last row of each table is the production selection, which is the configuration a player actually gets.

During the game session, inference should run on NPU or iGPU, which are unused during gaming sessions without driving significant resources. Even if run on the discrete GPU, it averages the 0.2% of resources. In real world usage, these compute consumptions are not impacting the gaming experience in meaningful ways.

Utilization tells us how much of the machine the features occupy. There is one more performance question to answer: whether each individual inference finishes quickly enough to stay inside the update rates described earlier.

The following measurements show the per-inference cost of each model on its production device, with a warm cache, on machines A to D.

All of these measurements in the tables above come from four configurations, taken in a benchmark host process with no game running alongside. Read them as the shape of the cost rather than as a guarantee for a given title on arbitrary hardware.

The WYVRN SDK allows developers to name events for game state, player state, abilities, actions, and environments. Those events can then trigger custom-designed Chroma lighting, haptic effects, or both at once. These events happen often and need to be handcrafted to generate the best experience.

Today those two layers are wired separately. The platform is a player-side setting in Razer Chroma, and the SDK is the developer's integration; That changes with the next phase of the WYVRN SDK, which brings the inference platform into the developer pipeline. A title will be able to enable and disable the three features from code, scene by scene, and combine them with authored effects at specific moments, so the adaptive layer becomes something you direct rather than something the player switches on. Timing and details will be announced through the WYVRN developer channels.

And taking that route does not mean building every effect from scratch. Razer Indie Lab offers a ready-made library of Chroma and haptic effects you can draw from, which covers a lot of ground, especially for small teams with no dedicated immersion budget. For larger productions, there is also a direct partnership route with Razer, where Razer’s immersive team can design effects for the game.

We are excited to continue pushing the edges of what haptics and RGB lighting can bring to the gaming immersion. More to come, stay tuned.