It’s based on the same event-based neuromorphic technology found on the earlier
AKD1000, relying on
spiking neural networks (SNN) to deliver real-time inference in a way that is much more efficient than traditional AI chips. What’s new here is that thanks to PCIe and SPI interfaces, the new ADK1500 co-processor can be paired to a wide range of hosts, ranging from Linux-capable application processors to resource-constrained microcontrollers with x86, Arm, or RISC-V architectures.
BrainChip AKD1500 key features and specifications:
- Akida Neuron Fabric
- Clocked at 5 to 400 MHz
- Delivering up to 800 effective GOPS at <1mW/GOPS
- On-device learning capabilities to enable “secure application personalization, without the need for a Cloud connection or retraining”
- On-Chip Conversion Complex
- Memory/Storage
- 1MB on-chip local memory
- SPI D/Q/O memory expansion interface (can also be used for sensors)
- Host interfaces
- PCIe Gen2 Endpoint Interface
- SPI S/D/Q/O Peripheral Interface
- Power dissipation – 250 mW typical at 400 MHz
- Dimensions – 7×7 mm MFCTFBGA169 package, 0.5 mm pitch
- Process – GlobalFoundries 22 nm FD-SOI CMOS digital logic process
BrainChip Software development platforms and workflow
The new Edge AI co-processor still relies on BrainChip’s MetaTF software flow, leveraging popular AI frameworks such as TensorFlow/Keras and ONNX/Pytorch. Developers can also make use of MCU SDKs like Edge Impulse or DeGirum, and compile and optimize their chosen models for the AKD1500. There’s some limited documentation (datasheet) available through
the developer website(free email registration required).
Potential applications for the AKD1500 include
personalized learning Edge AI systems,
Edge AI Vision Systems for ADAS/autonomous vehicles, tobots, drones, and video surveillance,
industrial IoTfor monitoring, control, and predictive maintenance, as well as
Smart Home devices such as Smart Speakers and other voice-controlled appliances.
BrainChip AKD1500 samples are available now, and volume production is scheduled for Q3 2026. Some early customers, namely Parsons, Bascom Hunter, and Onsor Technologies, have already designed AI-enabled sensing applications based on the AKD1500 co-processor for medical and defense-related applications.
BrainChip AKD1500 is an ultra-efficient Edge AI co-processor delivering up to 800 GOPS while consuming just 300 mW of power, making it suitable for
www.cnx-software.com