Cadence Pushes Wearables Forward with New Tensilica HiFi 1 DSP

New device improves user experience, boosts battery life and surpasses the HiFi 3 DSP

The new Cadence Tensilica HiFi 1 DSP. (Image courtesy of Cadence.)

The new Cadence Tensilica HiFi 1 DSP. (Image courtesy of Cadence.)

Cadence has introduced a new digital signal processor, the Tensilica HiFi 1 DSP. According to the company, it’s the smallest and lowest-power member of the HiFi DSP family, suited for small devices with battery limitations.

Rising consumer demand for high-performance wearable always-on devices with extended battery life is driving the electronics industry to advance innovations for earbuds, hearing aids, Bluetooth headsets, smartwatches, health trackers and other IoT devices. Cadence has responded by developing the Tensilica HiFi 1 DSP to offers better voice and music processing with optimal neural network capability within a small package using ultra-low energy.

With people increasingly tied to their tech devices wherever they go, there’s a greater need for compact forms that yield crisp audio and dependable connectivity that can last weeks between charges. For example, Xiaomi’s Redmi Smart Band Pro is a fitness wearable equipped with an array of sensors for workouts, music playback, weather forecasting and notifications for those on the move. It’s also powered by a 200mAh battery that the company states can last 14 days with typical use and 20 days in power-saving mode.

Cadence’s Tensilica HiFi 1 DSP is designed to take such always-on devices to the next level with its ultra-low energy consumption that extends the duration of voice communication and music playback for use in small form factor wearable and hearable devices. In addition, the product is also suited to automotive and industrial devices that require increased functionality coupled with low energy consumption.

“The HiFi 1 DSP from Cadence significantly lowers the energy required to run the always-on class of AI applications, such as the TensorFlow Lite Micro (TFLM) networks, speech wake word and person detect, enabling battery-constrained devices to run for longer lengths of time,” said Pete Warden, technical lead of TFLM at Google. “Cadence and Google have long collaborated on TFLM, and we are excited to continue the collaboration as Cadence pushes the boundaries of energy and performance further.”

Audio and visual codecs help minimize the data stream size before transmission while restoring the data post-transmission with high integrity. The Cadence Tensilica HiFi 1 DSP features a new LC3 codec, which provides high-resolution audio, forward error correction, low-delay modes and packet loss concealment.

The Tensilica HiFi 1 DSP provides ultra-low energy encoding and playback of LC3 and other Bluetooth codecs and ultra-energy keyword spotting for voice wake-up—all packaged within the small footprint HiFi DSP. The HiFi 1 DSP also offers an 11 to 16 percent lower area than the HiFi 3 DSP. Additionally, it has a 60 to 73 percent greater cycle and energy efficiency for ML-based “OK Google” keyword spotting and person-detect applications. The HiFi 1 DSP comparatively has over 18 percent greater cycle efficiency and 14 percent better energy efficiency for LC3 decoding.

“HiFi DSPs enjoy wide adoption in current-generation TWS [true wireless] earbuds and Bluetooth headsets,” said David Glasco, vice president of research and development for Tensilica IP at Cadence. “The advent of LC3 and wider market trends set the stage for next-generation hearables to offer a superior user experience and longer battery life. With many speech and voice algorithms migrating towards AI, we’re also seeing vastly expanding use cases for analytics and better sound quality in TWS earbuds. The compact HiFi 1 DSP enables these new use cases with ultra-low energy consumption, bringing to the mass market always-on and always-listening capabilities that were until now the privilege of premium products.”

The Tensilica HiFi 1 DSP’s features are manifold. Cadence states that it’s the most energy-efficient DSP for Bluetooth LE Audio due to its arithmetic coding for LC3 encoding and decoding. Talk time is extended through accelerated and highly efficient speech codecs. Machine learning reduces energy consumption, and memory access-optimized instruction set architecture yields better performance with small cache sizes. Efficient signal processing reduces energy and cycles for audio and speech, both pre-and post-processing. An optional, low-latency vector floating-point unit delivers higher floating-point throughput with lower energy consumption. Finally, the vector bool register improves energy efficiency for conditional code.