NDP120
Neural Decision Processor
Always-on Speech & Sensor-Fusion Processor
The Syntiant® NDP115 Neural Decision Processor™ is a special purpose chip for audio and sensor processing for always-on applications in battery powered devices and other power constrained systems. The NDP120 applies neural processing to run multiple applications simultaneously with minimal power consumption. Built using the Syntiant Core 2™ programmable deep learning architecture, NDP120 is designed to natively run multiple Deep Neural Networks (DNN) on a variety of architectures, such as CNN, RNN and fully connected networks. NDP120 brings a level of ML performance that delivers 25x the tensor throughput compared to the Syntiant Core 1™ embedded in Syntiant’s NDP100 and NDP101 devices. A programmable HiFi 3 DSP is available for classical audio processing.
The NDP115 supports dozens of application-defined audio sequences for a variety of use cases including:
| Far-field, near-field and close-talk voice interface
| Multiple wake words and local commands
| Acoustic Echo Cancellation (AEC), noise suppression, beamforming
| Speech enhancement
| Speaker identification and verification
| Acoustic event and scene classification
| Multi-sensor fusion
BLOCK DIAGRAM
KEY FEATURES & BENEFITS
| Hardware acceleration support for up to 6.4 GOPS/s
| Neural network layers supported: fully-connected, 2D convolution, depth-wise convolution, recurrent neural network including LSTM and GRU, average and max pooling
| Support for concurrent neural networks
| Up to 896k neural parameters in 8-bit mode, 1.8M parameters in 4-bit mode, and more than 7M parameters in 1-bit mode
| Quad PDM digital microphone interface
| Dual I2S channels or TDM4 streaming interfaces
| Support for up to 7 audio streams including I2S/ TDM output audio interface for streaming audio output, including post-processed audio
| I2C controller and target modes for sensor
control and integration
| QSPI target & controller interfaces
| 26 GPIO pins
| Programmable HiFi 3 DSP
| Input holding-tank with up to 10 seconds of audio recording and faster-then-real-time extraction
| Up to 100MHz internal operating frequency
| Embedded Arm Cortex-M0 for device management with 48KB SRAM, dual timers and UART functionality
| Low power PLL for flexible clock input
| Onboard firmware decryption and authentication
| Software Development Kit (SDK) integrates in any software environment
| Training Development Kit (TDK) to enable the use of standard frameworks such as TensorFlow for customer-programmed applications
| 3.1mm x 2.5mm 42-ball WLBGA package (0.4mm pitch)
| 5mm x 5mm 40-pin QFN package (0.4mm pitch) - also available as an AEC-Q100 Grade 3 qualified automotive SKU