High-Performance NPU

Overview

The ZIA™ A3000 AI processor IP is a low-power processor specifically designed for edge-side neural network inference processing. This versatile AI processor offers general-purpose DNN acceleration, empowering customers with the flexibility and configurability to optimize performance for their specific PPA targets. A3000 also supports high-precision inference, reducing CPU workload and memory bandwidth.

Key Features

High-Performance NPU: Featuring a scalable NPU architecture with over 40 TOPS of compute power, capable of handling a wide range of inference tasks.
Multicore Parallel Processing: A multicore design that can concurrently process multiple diverse models.
Mixed Precision Computation: Adopts a mixed precision computation approach to balance inference performance and accuracy.
Extensive Data Format Support: Supports a broad range of data formats including INT4/8, FP4/FP8/FP16.
Comprehensive Model Support: Wide coverage of ONNX operators, enabling a diverse set of AI models.
Edge Computing Optimization: Optimized for edge computing with high performance, power efficiency, and area efficiency(PPA).

Benefits

Scalable MAC Support: Each core supports from 96 to 2048 MAC units, enabling advanced multicore processing capabilities.
High-Performance Inference: Delivers over 40 TOPS of inference performance, executing complex computations at high speed.
Edge Device Optimization: Designed for low power consumption and cost, achieving high performance, power efficiency, and area efficiency (PPA). Compared to NPUs at the same level, it achieves around 50% smaller die size.
Mixed Precision Arithmetic: DMP proprietary mixed precision arithmetic unit enables combination of high accuracy and high-speed processing.
Broad Data Format Support: In addition to FP16 and INT8, the latest ML trends are supported, including INT4 and FP4, addressing diverse application needs.
Proprietary Profiler for Performance and Accuracy Analysis: a custom profiler providing per-layer performance analysis and accuracy degradation, facilitating the optimization of both speed and precision.

Block Diagram

Applications

Smart Cameras:
- Object detection, facial recognition, and other vision-based applications.
Industrial Automation:
- Predictive maintenance, quality inspection, and other machine vision tasks.
Robotics:
- Real-time obstacle detection and avoidance, autonomous navigation, and other intelligent control systems.
Internet of Things (IoT):
- Sensor data analysis, anomaly detection, and predictive maintenance for connected devices.

Deliverables

Datasheet
Hardware
- IP core specifications
- IP core data: Encrypted RTL
Software
- SDK/Tool specifications
- SDK/Tools
- Sample program

Technical Specifications

Maturity

Silicon Proven IP

Availability

NOW

Request Info