DSP Group Unveils DBM10 Low-Power Edge AI/ML SoC with Dedicated Neural Network Inference Processor

January 7 2021, 13:00
DSP Group announced the DBM10, a new low-power, cost-effective artificial intelligence (AI) and machine learning (ML) system-on-chip (SoC). This new open platform, with a cost- and power-optimized architecture, enables rapid development of AI and ML applications for mobile, wearables, hearables, and connected devices in general. It provides a complete platform in terms of voice and audio processing, without compromising the battery life of new designs, and allowing developers to implement their own differentiating algorithms.

DSP Group is a global provider of wireless and voice-processing chipset solutions with extensive experience in voice implementation and an increasing focus in advanced audio processing for personal audio with hearables (headphones, headsets, earbuds) and wearables (on-body electronics). Its new DBM10 SoC comprises a digital signal processor (DSP) and the company’s nNetLite neural network (NN) processor, both optimized for low-power voice and sensor processing in battery operated devices. 

This dual-core architecture offers developers with full flexibility of partitioning innovative algorithms between DSP and NN processor and enables fast time to market for integration of voice and sensing algorithms such as noise reduction, AEC, wake-word detection, voice activity detection and other ML models. 

The DBM10 features an open platform approach with a comprehensive software framework. This allows developers to quickly get next-generation designs to market with their own algorithms, or with DSP Group’s comprehensive and proven suite of optimized algorithms for voice, sound event detection (SED), and sensor fusion, as required by applications ranging from true wireless stereo (TWS) headsets to smartphones, tablets, wearables, and connected devices.

"Edge applications for AI are many and diverse, but almost all require the ultimate in terms of low power, small form factor, cost effectiveness, and fast time-to-market, so we are very excited about what the DBM10 brings to current and new customers and partners," says Ofer Elyakim, CEO of DSP Group. "Our team has worked to make the absolute best use of available processing power and memory for low-power AI and ML at the edge — including developing our own patent-pending weight compression scheme —while also emphasizing ease of deployment. We look forward to seeing how creatively developers apply the DBM10 platform."

The DBM10 adds to DSP Group’s SmartVoice line of SoCs and algorithms that are deployed globally in devices ranging from smartphones and laptops/PCs, to set-top boxes, tablets, remote controls, and smart IoT devices for the home. In 2020, SmartVoice shipments reached the 100 millionth milestone, and the new low-power DBM10 is already supported by an established ecosystem of third-party algorithm providers. Some of these have already begun running their NN algorithms on the nNetLite NN processor at the heart of the DBM10 to achieve maximum performance at the lowest power consumption.

Working alongside a programmable low-power DSP, the nNetLite processor supports all standard deep NN (DNN) and ML frameworks and employs a comprehensive cross-platform toolchain for model migration and optimization. The SoC device is supplied in a highly-compact form factor (~4 mm2), specified to support ultra-low-power inference at ~500 μW (typical) for voice NN algorithms and being able to run Hello Edge 30-word detection model @ 1 MHz (125 MHz available) as a reference. The DBM10 allows porting of large models (10s of megabytes) without significant accuracy loss using model optimization and compression.
related items