HSINCHU, Taiwan, Aug 10, 2026 – Andes Technology has launched the AndesAIRE AnDLA I370 v2.0 deep-learning accelerator IP and AndesAIRE NN SDK v1.2.0. The updates target edge systems running Transformer-based AI models.
The I370 v2.0 updates the company’s accelerator IP, while SDK v1.2.0 adds software tools for deploying Vision Transformers (ViT), Vision-Language Models (VLMs), and Small Language Models (SLMs).
Key I370 v2.0 specifications include:
The I370 v2.0 accelerator IP adds architectural changes to support multimodal vision-language model (VLM) and generative small language model (SLM) workloads:
- 32 to 2,048 MACs: Configurable MAC arrays for varying neural-network compute requirements.
- Up to 512-bit AXI bus width: Wider memory interfaces and additional DMA outstanding requests.
- Transformer architecture: A Programmable Tensor Operator Set Architecture for ViT, VLM and SLM workloads.
- Hardware operators: DIV and SQRT execution, plus operator fusion.
SDK Adds Attention Support and Linux and RTOS Components
AndesAIRE NN SDK v1.2.0 supports attention mechanisms used in Transformer models. For SLM inference, it separates prefill and decode phases to manage KV-cache placement and memory scheduling.
The SDK includes a kernel-mode driver and user-space runtime for Linux. It also supports RTOS environments with AnDLA task pre-emption for latency-sensitive workloads.
Model Advisor and Hardware Advisor map cost-performance Pareto fronts. The SDK v1.2.0 also includes word-level knowledge distillation to compress teacher models into student models for AnDLA deployment.
Other SDK updates include:
- LiteRT integration: Andes rebased AnDLA driver and NN Library wrapper functions on the latest LiteRT and LiteRT for microcontrollers AI frameworks.
- MLPerf Tiny v1.4 support: The SDK supports the MLCommons MLPerf Tiny v1.4 benchmark through LiteRT for microcontrollers integration.
- RTOS task pre-emption: AnDLA tasks can be pre-empted for real-time processing.
“Bringing Transformer models and VLMs to cost-sensitive edge devices is arguably one of the toughest engineering problems in the current AI landscape,” said Dr. Charlie Su, CTO and president of Andes Technology. “With the I370 v2.0 enhanced with a Programmable Tensor Operator Set Architecture and backed by our NN SDK v1.2.0, we have successfully unlocked on-device transformer models. Developers can now run ViT, VLM, and SLM locally under both RTOS and Linux, paving the way for a smarter generation of edge devices.”
Source: Andes Technology
About Andes Technology
Andes Technology is a supplier of high-performance, low-power 32- and 64-bit RISC-V CPU cores and related solutions for embedded systems. Founded in 2005, the company develops configurable processor IP cores used in applications such as AI, 5G communications, Internet of Things (IoT), storage, automotive electronics, and industrial control. Its portfolio includes RISC-V-based CPU cores, development tools, and software designed for energy-efficient, performance-optimized processors. Andes serves industries including consumer electronics, automotive, telecommunications, and industrial automation. Headquartered in Hsinchu Science Park, Taiwan, the company operates globally with a growing presence in Asia, North America, and Europe. Its customer base spans multiple regions, with Taiwan and the United States as key markets. The company continues to invest in R&D to advance processor technologies.
