Products Overview

Accelerator IP for
AI Designs

AI Accelerator IP That Scales from Edge Devices to Data Centers

The design of AI systems requires a careful balance of performance, power consumption, area (PPA), and latency. Origin™ Evolution™ NPU IP effectively manages the unique workload and memory requirements of applications that range from edge devices to automotive solutions and data centers. It is customizable for specific use cases and optimized to deliver the best PPA while handling various models, including large language models (LLMs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and others.

Optimized Architecture for Your Trained Networks

Origin Evolution offers unparalleled accuracy and predictable performance without requiring any hardware-specific optimizations or modifications to your trained model. Its patented packet-based execution architecture delivers single-core performance of up to 128 TFLOPS and maintains utilization rates of 70-90%. These metrics are measured on silicon while running typical AI workloads like Llama and YOLO. This outstanding performance and efficiency allow users to execute LLM, CNN, and other AI models with significantly lower power consumption compared to alternative solutions.

Silicon-proven and third-party tested, Expedera’s architecture is extremely power-efficient, achieving 18 TOPS/W. Origin Evolution NPUs offer deterministic performance, the smallest memory footprint, and are fully scalable. This makes them perfect for edge solutions with little or no DRAM bandwidth or high-performance applications such as autonomous driving. ASIL-B readiness-certified, Expedera offers the ideal solution for engineers looking for a single NPU architecture that easily runs LLM, CNN, RNN, and other network types while maintaining optimal processing performance, power, and area.

Full Software Stack

Expedera’s Origin Evolution IP platform includes a full TVM- and Relax-based software stack that enables developers to work efficiently as they deploy their networks on the target hardware. Supporting popular frameworks such as HuggingFace, Llama.cpp, PyTorch, Onnx, TensorFlow, TVM, and others, users can depoy their trained models as-is; no retraining or accuracy redutions are required.
The software stack automatically handles tasks such as packetization and debugging while providing user-friendly options such as mixed precision (integer and floating point) quantization—using your tools or ours—custom layer and custom network support, and multi-job APIs.

Origin Line of Products

Origin E1

The Origin E1 processing cores are individually optimized for a subset of neural networks commonly used in home appliances, edge nodes, and other small consumer devices. The E1 LittleNPU supports always-sensing cameras found in smartphones, smart doorbells, and security cameras.

Learn More

Products

Origin E1

Origin E1 neural engines are optimized for networks commonly used in always-on applications in home appliances, smartphones, and edge nodes that require about 1 TOPS performance. The E1 LittleNPU processors are further streamlined, making them ideal for the most cost- and area-sensitive applications.

Learn More Origin E1 Detail Link

Origin Evolution for Automotive​

Designed for performance-intensive automotive/ADAS applications, Origin Evolution for Automotive NPU IP cores excel at complex AI tasks, including computer vision, LLMs, warping, point cloud, grid sampling, image classification, and object detection. Single-core performance scales to 96 TFLOPS in a single core, with multi-core performance to PetaFLOPs.​

Learn More Origin E1 Detail Link

Download our White Papers

White Papers

Get in Touch With Us

Contact