GNA 5.0 is the ultra-low-power Gaussian and neural inference co-processor dedicated to continuous audio and speech workloads. The intelligence backbone of the Klangkerne audio stack—always-on, always-listening, always-aware.
Gaussian Mixture Model (GMM) co-processors have been integral to Intel silicon since Ice Lake (GNA 1.0) through Arrow Lake (GNA 3.5). As the NPU gained prominence, GNA appeared to be deprecated—until Eventide's audio-first architecture revived the concept.
Where GNA 3.5 focused purely on audio inference with INT8 precision, GNA 5.0 expands the charter:
| Attribute | Details |
|---|---|
| Precision Support | INT8, BF16 |
| Operation Types | GMM (Gaussian Mixture Model), 1D Convolution, 2D Convolution |
| Max Power (Active) | 45 mW (sustained audio processing) |
| Idle Power | <1 mW (always-on standby) |
| Process Node | Intel 14A-E (enhanced for minimum leakage) |
| Integration | Klangkerne tile (direct coupling on uncore) |
| OpenVINO Support | OpenVINO 5.0 GNA Plugin (optimized kernel library) |
| Key Use Cases | Noise cancellation, wake-word detection, speaker ID, acoustic context |
Real-time suppression of background noise, wind, and environmental interference— powered by trained DNN models executing on GNA 5.0 without CPU overhead.
Always-listening for custom wake phrases. GNA 5.0 monitors microphone input continuously, waking the CPU only when activated—critical for low-power always-on assistants.
Identify speakers and separate concurrent voices in multi-party conversations. Audio-visual context awareness enables personalization without high power cost.
Classify audio environment (office, car, outdoors, music playback) to tune processing parameters dynamically—all on-device, in real-time, autonomously.