Syntiant's Third-Gen "Neural Decision" Chip, the NDP250, Packs 30 GOPS in a Microwatt Envelope
Five times faster than its Core 2-powered predecessors, the NDP250 is being positioned for always-on voice and vision applications.
On-device deep learning specialist Syntiant has announced its third-generation "Neural Decision Processor," the NDP250 — delivering a claimed 30 giga-operations per second (GOPS) of compute for everything from speech synthesis to image recognition within a microwatt power envelope.
"Our NDP250 builds on two generations of neural network architectures to deliver 30 GOPS, making it our fastest, highest-performing chip yet," says Syntiant chief executive officer Kurt Busch of the new chip, unveiled at Embedded World today.
"Compatible with a host of architectures while running multiple different layers simultaneously at significantly less power than existing solutions," Busch continues, "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."
The NDP250 uses the same approach as the company's previews "Neural Decision Processors," but with the new third-generation Syntiant Core 3 architecture — delivering, the company says, five times the performance of the company's second-generation parts without ballooning the power requirements. The chip includes an Arm Cortex-M0 processor in addition to Syntiant's Core 3 and a HiFi3 programmable digital signal processor (DSP), along with a dual 11-wire direct image interface, dual PDM digital microphone interface, I2S with PCM, and quad-SPI and dual-I2C buses along with 58 general-purpose input/output (GPIO) pins.
According to Syntiant, the new chip can handle multiple heterogeneous networks running simultaneously with over six million total neural parameters when running in eight-bit precision mode, 1D, 2D, and depthwise convolutional neural networks (CNNs), fully-connected networks, and recurrent neural networks including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. As well as real-time speech recognition and voice synthesis tasks, the company says the chip has the power to run always-on image recognition models with a power draw below 30mW.
The NDP250 is being demonstrated at Syntiant's booth at Embedded World 2024 this week, Hall 2 Booth #2-238; the company has confirmed it is sampling the part now, but has not offered a schedule for general availability.
More information on Syntiant's chips is available on the company's website.