Syntiant unveils NDP250 neural decision processor

Built using the Syntiant Core 3 programmable deep learning architecture, the NDP250 is designed for numerous imaging, speech, and sensor applications in power-constrained devices, such as person detection, object classification, automatic speech recognition (ASR), text to speech (TTS), and motion tracking, among others.

“Our NDP250 builds on two generations of neural network architectures to deliver 30 GOPS, making it our fastest, highest-performing chip yet,” said Kurt Busch, CEO of Syntiant. “Compatible with a host of architectures while running multiple different layers simultaneously at significantly less power than existing solutions, the NDP250 with our new Core 3 engine is the ideal real-time speech interface for large language models, and can bring powerful AI to battery-powered, always-on vision applications in automotive security, appliances, cameras, smart displays, and video doorbells. With the NDP250, applications that previously required power that was measured in watts can now be done with power measured in microwatts.”

A faster interface for large language models (LLM)

Integrating ASR and TTS directly into devices running on the NDP250 creates a smooth pathway to utilising LLMs as the next generation voice interface. The chip’s advanced processing capabilities enable these features to understand and produce human language more accurately and fluently, improving user interactions and experiences.

“Syntiant’s third generation of at-memory compute processors for multi-modal artificial intelligence from a variety of sensor types extends their lead into system designs, where LLMs will interface with humans and the environment at minimal latency and energy cost,” said Michael Stewart, Managing Partner at M12, Microsoft’s Venture Fund.

Moving vision AI from the Cloud to the device

The NDP250 has advanced image capabilities and is ideal for ultra-low power video applications for automotive security, battery-powered cameras, and video doorbells.

Running powerful always-on image recognition at under 30mW has several advantages, including:

• A significant reduction in power consumption by processing data locally on devices, thereby extending battery life and enabling efficient resource utilisation.
• Lower latency since data doesn't need to travel back and forth to a remote server, resulting in faster response times crucial for real-time applications and greatly increases customer satisfaction.
• Enhanced privacy by processing sensitive data locally, minimising the need to transmit information over networks where it could be vulnerable to breaches or interception.
• A notable reduction in cloud costs, sometimes as high as 90%, since less data needs to be transferred and processed in the cloud, leading to lower infrastructure expenses for businesses deploying edge AI solutions.

Equipped with an Arm Cortex M0 processor and a Hi-Fi 3 DSP to support feature extraction and signal processing for image and voice enhancements, the NDP250’s integrated power management unit allows single power rail operation, where the integrated phase-locked Loop (PLL) provides further system cost and size optimisation.

With the ability to process multiple heterogenous networks concurrently, the NDP250 also supports convolution neural networks including 1D, 2D and depth-wise, fully connected networks, and recurrent neural networks including LSTM (long short-term memory) and GRU (gated recurrent unit).

Other key features include: 

• Syntiant Core 3 neural network
• Supports more than 6M neural parameters (in 8-bit mode)
• Hardware acceleration over 30 GOPS 
• Dual 11-wire direct image interface
• Dual PDM digital microphone interface
• I2S serial interface with PCM
• Quad-SPI and dual-I2C controller and target for multi-modal sensor fusion
• Up to 120MHz internal operating frequency
• Low power PLL for flexible clock input
• Software development kit
• Training development kit
• 120-ball 6.1 x 5.1mm eWLB package (0.5mm pitch)

The NDP250 is sampling now.

Syntiant will be demonstrating the NDP250 and various hardware-agnostic deep learning vison models at embedded world 2024 (Hall 2 – Booth 2-238), 9–11 April 2024 in Nuremberg, Germany.