NDP120
The Syntiant NDP120 is a second-generation Neural Decision Processor designed for always-on, battery-powered edge AI applications such as keyword spotting, smart sensors, and low-power audio recognition.
Publisher review
The Syntiant NDP120 is a second-generation Neural Decision Processor designed for always-on, battery-powered edge AI applications such as keyword spotting, smart sensors, and low-power audio recognition. It targets developers who need to run multiple deep neural networks simultaneously on a tiny energy budget, from earbuds to automotive systems. Built on the Syntiant Core 2 programmable deep learning architecture, the NDP120 natively supports CNNs, RNNs, and fully connected networks, and integrates a DSP for flexible pre/post-processing.
In the MLPerf Tiny v1.3 benchmark suite (September 2025), it led the new streaming wake word detection test, achieving both the lowest energy per inference and highest throughput, with a streaming workload using only a ~2% duty cycle at 50 MHz. Compared to traditional microcontrollers, the NDP120 delivers 10–100x better energy efficiency, making it suitable for always-on applications that previously required trade-offs between responsiveness and battery life. The device can run multiple applications concurrently — such as noise cancellation, beamforming, or sensor fusion — with minimal additional power draw, as demonstrated by its ability to process continuous audio input efficiently in a realistic deployment test.
The NDP120 offers 25x tensor throughput compared to earlier Syntiant NDP devices, and its toolchain provides a straightforward path from reference or customer models to deployment, with compatibility for common layers and flexible pre/post-processing on the integrated DSP. Syntiant, founded in 2017 and headquartered in Irvine, Calif., has deployed over 100 million purpose-built silicon and ML models, along with billions of MEMS microphones and sensors. Competitors include Latent AI, SambaNova, Mythic, Unconventional AI, Applied Brain Research, and Kneron, but the NDP120's specific advantage lies in its proven benchmark leadership for streaming audio workloads and its ability to run multiple neural network types without extensive post-training compilation. Honest trade-offs include limited public pricing information, reported issues with firmware updates on Windows Command Prompt, a focus primarily on audio and sensor applications rather than general-purpose edge AI, and reliance on a proprietary architecture that may require vendor-specific toolchain adoption.
How it works
-
Streaming wake word detection
Led the new MLPerf Tiny v1.3 streaming benchmark with lowest energy per inference and highest throughput, using ~2% duty cycle at 50 MHz.
-
Multi-network support
Natively runs CNNs, RNNs, and fully connected networks simultaneously, enabling concurrent tasks like noise cancellation and beamforming.
-
10-100x energy efficiency
Compared to traditional microcontrollers, delivers 10-100x better energy efficiency for always-on audio and sensor applications.
-
25x tensor throughput
Offers 25x tensor throughput versus earlier Syntiant NDP devices, without requiring extensive post-training compilation.
-
Integrated DSP flexibility
Includes an integrated DSP for flexible pre/post-processing, with typical networks using only a fraction of on-chip resources.
-
Low-power concurrent execution
Can run multiple applications simultaneously with minimal battery power consumption, as shown by the ~2% duty cycle at 50 MHz.
-
Proven benchmark performance
Demonstrated strong performances across multiple MLPerf Tiny benchmark rounds, including both static and streaming tests.
Strengths and trade-offs
Strengths
- Achieved both the lowest energy per inference and highest throughput in the MLPerf Tiny v1.3 streaming wake word detection benchmark.
- Delivers 10-100x better energy efficiency compared to traditional microcontrollers, enabling always-on battery-powered applications.
- Offers 25x tensor throughput compared to earlier Syntiant NDP devices without requiring extensive post-training compilation.
- Supports multiple neural network architectures (CNNs, RNNs, fully connected) and flexible pre/post-processing on an integrated DSP.
Trade-offs
- Pricing is not publicly specified, making cost comparison with competitors difficult for potential buyers.
- Users have reported issues with firmware updates when using Windows Command Prompt, which may complicate deployment workflows.
- The processor is optimized primarily for audio and sensor applications, limiting its suitability for general-purpose edge AI tasks.
- Relies on a proprietary Syntiant Core 2 architecture, requiring developers to adopt the vendor-specific toolchain for model deployment.
Pricing context
Not publicly specified in the available sources.
Getting started with NDP120
-
Sign up for Syntiant account
Go to the Syntiant developer portal and create an account. Provide your email and set a password. Verify your email to activate the account and gain access to the NDP120 toolchain and documentation.
-
Install the NDP120 toolchain
Download and install the Syntiant toolchain for your operating system. Follow the installation guide to set up the compiler, debugger, and deployment utilities. Ensure your system meets the prerequisites listed in the documentation.
-
Connect the NDP120 evaluation board
Connect the NDP120 evaluation board to your computer via USB. Install any required drivers. Verify the board is recognized by the toolchain using the provided diagnostic commands.
-
Load a pre-trained audio model
Download a pre-trained keyword spotting model from the Syntiant model zoo. Use the toolchain to compile the model for the NDP120. Transfer the compiled model to the evaluation board using the deployment utility.
-
Run the model and verify output
Power the board with a battery or USB. Speak the target keyword into the onboard microphone. Observe the output indicator or serial console to confirm the model detects the keyword with low latency.
Frequently Asked Questions
What is the Syntiant NDP120 and what is it used for?
The Syntiant NDP120 is a second-generation Neural Decision Processor for always-on, battery-powered edge AI. It is used for keyword spotting, smart sensors, and low-power audio recognition, running multiple deep neural networks simultaneously on a tiny energy budget in devices like earbuds.
How does the NDP120 perform in MLPerf Tiny v1.3 benchmarks?
In the MLPerf Tiny v1.3 streaming wake word detection test, the NDP120 led with the lowest energy per inference and highest throughput. It achieved this using only a ~2% duty cycle at 50 MHz, demonstrating exceptional efficiency for streaming audio workloads.
What energy efficiency does the NDP120 offer compared to microcontrollers?
The NDP120 delivers 10 to 100 times better energy efficiency than traditional microcontrollers. This makes it suitable for always-on applications that previously required trade-offs between responsiveness and battery life, enabling continuous audio processing with minimal power draw.
What neural network types does the NDP120 support?
The NDP120 natively supports CNNs, RNNs, and fully connected networks simultaneously. It also integrates a DSP for flexible pre/post-processing, allowing concurrent tasks like noise cancellation and beamforming without extensive post-training compilation.
What are the main weaknesses of the NDP120?
Pricing is not publicly specified, complicating cost comparisons. Users report firmware update issues on Windows Command Prompt. The processor focuses on audio and sensor applications, not general-purpose edge AI. It also requires adopting Syntiant's proprietary toolchain for model deployment.
How does the NDP120 compare to earlier Syntiant processors?
The NDP120 offers 25 times higher tensor throughput compared to earlier Syntiant NDP devices. This improvement comes without requiring extensive post-training compilation, enabling faster and more efficient processing for multiple neural networks in battery-powered edge AI applications.
Alternatives
How NDP120 compares
Direct head-to-head against 2 competitors. Picked by 7wData.
NDP120
- Pricing
- Not publicly specified in the available sources.
- Target
- The Syntiant NDP120 is a second-generation Neural Decision Processor designed for always-on, battery-powered edge AI applications such as keyword spotting, smart sensors, and low-power audio
- Strength
- Achieved both the lowest energy per inference and highest throughput in the MLPerf Tiny v1.3 streaming wake word detection benchmark.
- Watch for
- Pricing is not publicly specified, making cost comparison with competitors difficult for potential buyers.
Arduino Nicla Voice
- Pricing
- $49/unit
- Target
- Developers prototyping voice AI
- Deployment
- Embedded
- Strength
- Arduino ecosystem integration
- Watch for
- Limited neural network capacity
Syntiant NDP101
- Pricing
- Custom/Contact sales
- Target
- Low-power edge voice applications
- Deployment
- Embedded
- Strength
- Ultra-low power consumption
- Watch for
- Requires Syntiant toolchain
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.