Actions Semiconductor's ATW6095 on-device AI SoC enters mass production, empowering high-end AI Note

The hardware competition in AI Notes is shifting from "being able to record" toward "hearing more clearly, computing more efficiently, and running longer."

Actions Technology has continued to deepen its work in the AI Note field. Its ATS308X chip series has empowered the Honor Choice KUMI AI Note, the Mobvoi TicNote (WeCom Edition), the Memoket Gem, the Xueersi Notes xNote, and many other mainstream products at home and abroad to land in office and education scenarios, earning broad market recognition.

Recently, a number of high-end AI Notes powered by Actions' new-generation edge AI SoC chip, the ATW6095 series, have officially entered mass production and reached the market. The end products include the Yidao Intelligent YoooClaw C·ONE Pro and the Xinren Intelligent N.B.BOSS P2 Plus. Relying on a self-developed compute-in-memory NPU, high-quality multi-microphone pickup, and a system-level low-power design, the ATW6095 series helps brand customers balance pickup in complex scenarios, local AI processing, and continuous recording within a limited battery budget, helping high-end AI Notes create differentiated, benchmark-level product experiences.

Computing efficiently: a compute-in-memory NPU that lets local AI run efficiently

For device makers, high-end AI Notes have long been constrained by the mutual tug-of-war between on-device AI capability and device battery life: carrying more complex local AI models noticeably reduces continuous working time, while guaranteeing long working hours requires lowering local AI processing specifications. The change brought by the Actions® ATW6095 chip series lies precisely in resolving this trade-off dilemma.

The chip is equipped with Actions Technology's self-developed compute-in-memory NPU, which moves the computing units inside the memory array, reducing the huge energy overhead of repeatedly moving data between storage and computing units under traditional architectures.

Thanks to this, the NPU in the ATW6095 series delivers 100GOPS of single-core computing power, with an energy efficiency ratio of up to 6.4TOPS/W at INT8 precision. Compared with mainstream DSPs and CPUs under traditional architectures, it offers an energy efficiency advantage of more than tenfold to several dozenfold. This means that when running the same AI inference tasks, the ATW6095 series consumes less power; under the same power budget, it can carry higher-specification AI algorithm models.

For complex AI computations that traditional DSPs struggle to support and that require sustained computing power of more than 10GOPS, Actions Technology works in deep collaboration with Aizip, a leading edge AI algorithm team from Silicon Valley. Leveraging the powerful and efficient compute-in-memory NPU in the ATW6095 series, the solution natively adapts Aizip's industry-leading lightweight far-field voice enhancement and sound source separation models.

The solution supports far-field pickup at distances of up to 10 meters; in complex scenarios with multiple types of noise, the ASR character error rate/word error rate can be reduced by up to 60% relatively; in directional pickup mode, it can achieve up to 40dB of interfering human voice suppression, effectively improving speech recognition accuracy. In the future, the solution can also support multiple people speaking simultaneously based on an omnidirectional multi-microphone array, separating mixed speech in real time into 2–3 independent voice signals, further reducing the character error rate/word error rate, providing high-quality voice input for cloud ASR and language models, and supporting richer and more challenging AI application scenarios.

The ATW6095 series also integrates a 420MHz HiFi5 DSP, which improves audio processing performance by 200% and neural network processing performance by 400% compared with traditional HiFi4 products.

The HiFi5 DSP and the compute-in-memory NPU work in deep synergy, breaking the limitations of simple hardware stacking: latency-sensitive front-end pickup processing is handled by the DSP, while the NPU focuses on AI inference; for computing tasks that the NPU does not yet directly adapt but that have already completed software debugging, the HiFi5 DSP can also take on supplementary computation. The two divide work efficiently and fully support core algorithms such as beamforming, bone-microphone fusion, AGC, voice enhancement, noise reduction, and directional pickup, unleashing full computing power within limited power consumption.

Hearing clearly: a high-specification audio chain that supports pickup in complex scenarios

A powerful AI core and algorithms also require high-quality raw audio signal input. The ATW6095 series adopts a single-chip highly integrated design that coordinates audio perception with edge computing, forming a complete closed loop of capabilities.

The chip integrates Actions' self-developed high-quality ADC/DAC modules, with ADC SNR reaching 112dB, THD+N as low as -100dB, and DAC SNR reaching 113dB. The single-chip audio performance reaches the level of international first-tier standalone audio codecs, ensuring the output of raw audio signals with low noise floor, low distortion, and high fidelity, building a solid foundation for accurate AI recognition.

In terms of microphone interfaces, the ATW6095 series supports 4 channels of AMIC + 8 channels of DMIC, higher than conventional industry specifications, providing more ample interface options for multi-microphone array designs. Brand customers can flexibly plan the number and deployment positions of microphones according to product usage scenarios. With multi-microphone array designs, they can comfortably handle complex usage scenarios such as noisy environments, long distances, and multiple people talking at once.

Recording for long: system-level low power supporting long continuous recording

For AI Notes that need to work for long periods, battery life is a key factor affecting user experience. Measured results show that when equipped with 4+1 microphones and running multiple types of edge AI algorithms, the ATW6095 series can keep the average working current of the whole device during recording within 10mA, saving more than 20% of power consumption compared with other solutions at the same working voltage.

The value brought by low power consumption is not only longer battery life, but also more room for hardware capabilities.

The ATW6095 series integrates a 2D graphics acceleration engine and supports mainstream GUI frameworks such as LVGL, enabling local display of recording status, battery level, and audio waveforms. Graphics processing capability works in concert with the low-power design, offering more options for product forms that balance intuitive interaction and long-duration recording.

The chip requirements of high-end AI Notes concentrate a comprehensive test of sound acquisition, real-time signal processing, neural network inference, and system power management.

The ATW6095 series is positioned directly at the high-end AI hardware market, integrating the self-developed compute-in-memory NPU, HiFi5 DSP, high-specification audio chain, and system-level low-power design into a single chip. The application boundaries of this capability combination go far beyond AI Notes. AI glasses, AI desktop assistant robots, and a broader range of AI portable smart hardware are all scenario areas to which the ATW6095 series is highly suited.