NDP115
The Syntiant NDP115 Neural Decision Processor is a special-purpose chip designed for always-on deep learning inference in battery-powered and power-constrained edge devices.
Publisher review
The Syntiant NDP115 Neural Decision Processor is a special-purpose chip designed for always-on deep learning inference in battery-powered and power-constrained edge devices. Built using the Syntiant Core 2 programmable architecture, it delivers 25x the tensor throughput of the earlier Core 1 generation found in the NDP100 and NDP101. The NDP115 targets applications in mobile phones, smart home IoT endpoints, hearables, wearables, AR/VR, and industrial IoT, enabling cloud-free audio and sensor AI processing in form factors as small as 2.1 mm x 2.1 mm. It is now shipping in production volumes and is priced at $3.25 per unit in 10,000-unit quantities, making it a cost-effective option for high-volume consumer and commercial products.
The NDP115 natively runs multiple deep neural networks concurrently on architectures including CNNs, RNNs, and fully connected networks, supporting up to five audio streams simultaneously. It achieves speech inferences at just 280 microwatts, enabling always-on voice interfaces without draining batteries. The chip integrates a programmable HiFi-3 DSP with 144KB of 64-bit data RAM and 64KB of 64-bit instruction RAM for flexible feature extraction and classical audio processing, along with an Arm Cortex M0 microcontroller for device management. It supports direct PCM audio input, I2C controller and target modes for sensor integration, pulse density modulation (PDM) interfaces, up to 13 GPIO pins, flexible clock generation, and onboard firmware decryption and authentication.
Compared to general-purpose MCUs or DSPs, the NDP115 offers a dedicated neural processing path that offloads inference from the main application processor, reducing system power and latency. Competitors such as the Ambarella CV22 or the Hailo-8L target similar edge AI use cases but typically require higher power budgets and larger packages. The NDP115's 25-ball WLBGA package (0.4mm pitch) and sub-milliwatt inference capability give it a distinct advantage in ultra-compact, battery-operated devices where size and power are critical constraints. Syntiant's proprietary Core 2 engine also provides a higher level of integration for multi-sensor fusion than many competing neural accelerators.
The primary trade-off is that the NDP115 is a fixed-function neural processor, not a general-purpose compute platform; developers must work within Syntiant's toolchain and model zoo, which limits flexibility for custom algorithms. While the chip supports multiple network architectures, its performance on very large or complex models may be constrained by the 144KB data RAM and 64KB instruction RAM of the DSP. The package's 0.4mm ball pitch requires advanced PCB assembly techniques, potentially increasing manufacturing complexity. Additionally, detailed benchmark comparisons against competing edge AI processors are not publicly available from Syntiant, making independent performance validation difficult for prospective buyers.
How it works
-
Syntiant Core 2 neural processor
Programmable deep learning architecture that delivers 25x the tensor throughput of the prior Core 1 generation.
-
Concurrent multi-network support
Runs multiple DNNs simultaneously on architectures like CNN, RNN, and fully connected networks.
-
HiFi-3 DSP with 144KB data RAM
Embedded user-programmable DSP with 144KB 64-bit data RAM and 64KB 64-bit instruction RAM for audio processing.
-
Ultra-low power speech inference
Runs speech inferences at only 280 microwatts, enabling always-on voice detection in battery devices.
-
Multi-sensor fusion interfaces
Supports I2C and PDM interfaces for sensor fusion including accelerometers, magnetometers, and pressure sensors.
-
Compact 25-ball WLBGA package
Measures 2.1 mm x 2.1 mm with 0.4mm ball pitch, suitable for ultra-compact edge devices.
-
Onboard firmware security
Includes firmware decryption and authentication to protect deployed neural network models.
Strengths and trade-offs
Strengths
- Achieves speech inference at 280 microwatts, enabling always-on voice processing without significant battery drain.
- Delivers 25x the tensor throughput of Syntiant's Core 1 architecture, improving performance for multi-network loads.
- Supports concurrent execution of multiple neural network architectures including CNNs, RNNs, and fully connected networks.
- Priced at $3.25 per unit in 10,000-unit quantities, offering a cost-effective solution for high-volume edge AI products.
Trade-offs
- Limited to Syntiant's proprietary toolchain and model zoo, restricting flexibility for custom or non-standard algorithms.
- DSP memory is constrained to 144KB data RAM and 64KB instruction RAM, which may limit complex model sizes.
- The 0.4mm ball pitch package requires advanced PCB assembly, potentially increasing manufacturing costs.
- Independent performance benchmarks against competing edge AI processors are not publicly available from Syntiant.
Pricing context
$3.25 per unit in 10,000-unit quantities.
Getting started with NDP115
-
Purchase NDP115 samples
Order the NDP115 chip from Syntiant's sales channel or an authorized distributor. For prototyping, request evaluation kits that include the chip on a breakout board to simplify initial testing and development.
-
Set up Syntiant toolchain
Download and install Syntiant's development tools from their official website. Register for a developer account to access the model zoo, documentation, and firmware libraries needed to compile and deploy neural networks on the NDP115.
-
Connect NDP115 to host
Wire the NDP115 evaluation board to your host microcontroller or single-board computer using I2C or SPI interfaces. Connect a PDM microphone for audio input and verify power supply within the chip's specified voltage range.
-
Load a pre-trained model
Select a keyword spotting or audio classification model from Syntiant's model zoo. Use the provided tools to compile the model into firmware, then flash it onto the NDP115 via the I2C bootloader interface.
-
Test always-on inference
Power the NDP115 and speak a trigger phrase into the microphone. Monitor the chip's output pin or I2C register for inference results. Measure power consumption to confirm sub-milliwatt operation during always-on listening.
Frequently Asked Questions
What is the Syntiant NDP115 and what is it used for?
The Syntiant NDP115 is a special-purpose chip for always-on deep learning inference in battery-powered edge devices. It processes audio and sensor AI without a cloud connection, targeting applications like mobile phones, hearables, wearables, and industrial IoT.
How much power does the NDP115 consume for speech inference?
The NDP115 achieves speech inferences at just 280 microwatts, enabling always-on voice interfaces without draining batteries. This ultra-low power consumption makes it ideal for compact, battery-operated devices where energy efficiency is critical.
What is the price of the Syntiant NDP115?
The NDP115 is priced at $3.25 per unit in 10,000-unit quantities, making it a cost-effective option for high-volume consumer and commercial products. It is now shipping in production volumes.
How does the NDP115 compare to other edge AI processors like the Ambarella CV22?
Competitors like the Ambarella CV22 typically require higher power budgets and larger packages. The NDP115's sub-milliwatt inference and 2.1 mm x 2.1 mm package give it an advantage in ultra-compact, battery-operated devices where size and power are critical.
What are the main limitations of the NDP115?
The NDP115 is a fixed-function neural processor limited to Syntiant's toolchain and model zoo, restricting custom algorithms. Its DSP has 144KB data RAM and 64KB instruction RAM, which may constrain large models. The 0.4mm ball pitch also requires advanced PCB assembly.
What neural network architectures does the NDP115 support?
The NDP115 natively runs multiple deep neural networks concurrently, supporting architectures including CNNs, RNNs, and fully connected networks. It can handle up to five audio streams simultaneously, enabling complex multi-network inference in edge devices.
Alternatives
How NDP115 compares
Direct head-to-head against 2 competitors. Picked by 7wData.
NDP115
- Pricing
- $3.25 per unit in 10,000-unit quantities.
- Target
- The Syntiant NDP115 Neural Decision Processor is a special-purpose chip designed for always-on deep learning inference in battery-powered and power-constrained edge devices.
- Strength
- Achieves speech inference at 280 microwatts, enabling always-on voice processing without significant battery drain.
- Watch for
- Limited to Syntiant's proprietary toolchain and model zoo, restricting flexibility for custom or non-standard algorithms.
Cadence Tensilica HiFi DSP
- Pricing
- Custom/Contact sales
- Target
- Audio processing in wearables
- Deployment
- Embedded
- Strength
- Optimized for audio DSP workloads
- Watch for
- Complex licensing structure
CEVA SensPro2
- Pricing
- Custom/Contact sales
- Target
- Sensor fusion in IoT devices
- Deployment
- Embedded
- Strength
- High-efficiency sensor processing
- Watch for
- Limited audio-specific optimizations
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.