The company’s latest Deep Learning Accelerator (DLA) it has been engineered to bring advanced neural network performance to cost-sensitive edge and endpoint AI applications.
Building on its predecessor, the AnDLA I350, the new AnDLA I370 significantly broadens AI inference capabilities by adding support for Audio/Voice AI applications, RNN models, INT16 data type, and configurable bus widths. It has been tailored for performance, scalability, and power efficiency and been designed to meet the growing demands for on-device intelligence in edge computing, IoT, and embedded vision applications.
The AndesAIRE AnDLA I370 supports industry-standard deep learning frameworks including TensorFlow Lite, PyTorch, and ONNX, so developers can easily deploy AI workloads across platforms. Capable of executing complex neural network operations such as convolution, fully connected, elementwise, pooling, activation, channel padding and up-sampling, the I370 integrates internal DMA and local memory and boosts execution efficiency through operator and layer fusion.
Key features include:
- Up to 2 TOPS (Tera Operations Per Second)/GHz
- Configurable MAC count and local memory size for usage flexibility
- Expanded model support for recurrent neural networks (RNNs)
- Enhanced INT16 and INT8 precision support for optimal performance-efficiency balance
Alongside the new accelerator, Andes is also offering the AndesAIRE NN SDK, a comprehensive toolkit for end-to-end neural network development. Built to accelerate deployment on AnDLA and AndesCore RISC-V platforms, the SDK includes:
- AndesAIRE NNPilot – Model analysis, pruning, quantization, and deployment tool suite
- LiteRT/LiteRT for Microcontrollers (formerly TFL/TFLM) –Inference frameworks optimised for various host environments
- AnDLA Driver and Runtime – Auto-generated code for efficient integration with applications
- AndesAIRE XNNPACK – Andes-optimised high-performance AI compute library based on open source
NNPilot streamlines the process from models to deployment by automatically converting models into optimised command images, completing with sample host code to simplify integration in bare-metal or embedded environments.
With the AnDLA I370 and the AndesAIRE NN SDK, Andes said that it was helping to “pave the way for flexible, high-efficiency AI acceleration at the edge.”
“AI at the edge demands compact yet powerful computing solutions,” said Dr. Charlie Su, CTO and President of Andes Technology. “The AnDLA I370 delivers flexibility, high throughput, and seamless integration with popular frameworks, empowering developers to bring intelligence to a broader range of devices.”
“What sets Andes apart is the seamless fusion of our leading AndesCore RISC-V processors with our purpose-built AnDLA deep learning accelerators,” said Simon TC Wang, Senior Technical Marketing Manager of Andes Technology. “RISC-V processors and NN Library secure the flexibility and future-proof for the rapid growth of the AI/ML applications while the accelerator guarantees the high performance and the compute density. This tight integration leverages the best of both worlds to enable a highly efficient, scalable, and configurable AI/ML platform that is uniquely positioned to meet the demands of edge and endpoint intelligence.”
