Anmou Technology officially announced its self-researched new generation of artificial intelligence processor “Cortex X2 NPU. The processor not only provides significant improvements in arithmetic power, accuracy and flexibility, but also provides special optimization for application scenarios such as in-vehicle and edge computing.
With the launch of Cortex X2 NPU, Anmou Technology released Cortex NPU software open source program to meet customers’ needs for more autonomous and flexible algorithm transplantation through open source code.
Nowadays, with the booming development of smart car industry and edge computing, the demand for AI algorithms in various scenarios has increased exponentially, whether from 720P to 4K resolution, or from single-channel image to multi-channel image fusion analysis.
As a new generation AI processor, the Cortex X2 NPU adopts the third-generation Cortex architecture and supports multi-core Cluster with up to 320TOPS subsystems. Real-time hardware task management allows the X2 NPU to achieve up to 10 million times/second task scheduling, optimizing the performance of each computing unit. While the computing power is greatly increased, the Cortex X2 NPU also has higher accuracy and flexibility. In terms of precision, Cortex X2 NPU supports int4 / int8 / int12 / int16 / int32, fp16 / bf16 / fp32 multi-precision fusion computing, which significantly improves the computational efficiency and density. In terms of flexibility, Cortex X2 NPU supports custom operators to meet the deployment needs of various models, and also provides customized AI solutions for various application scenarios to further meet the differentiated needs of customers in scenarios such as smart driving, cell phone image AI processing, and human-computer interaction.
The Cortex X2 NPU has a lot of performance optimizations for ADAS (Advanced Driver Assistance System), intelligent cockpit, tablet, desktop and cell phone applications. It can significantly improve the performance of applications such as high-resolution image processing in cell phone photography and video recording, as well as Transformer (Transformer is a transformation model that relies entirely on a self-attentive mechanism to calculate its input and output representation) commonly used in vehicles. The i-Tiling technology also significantly reduces bandwidth requirements and further improves computational efficiency, allowing customers to more easily cope with the diverse computational demands of constant iteration.
To help developers port and debug algorithms more easily and quickly, the Cortex X2 NPU also provides a comprehensive AI software platform that can better meet developers’ needs for performance tuning and system deployment. Currently, the Cortex X2 NPU has been officially delivered to customers, and there will be a number of chips equipped with the Cortex X2 NPU available this year.
Under the Cortex NPU software open source program, Anmoo has taken the lead in opening up the NPU intermediate representation layer specification, model parser, model optimizer, driver, etc., and provided the “Cortex ” Compass software platform to relevant partners, including software simulators, debuggers, C compilers and other software tools. In order to meet the needs of partners for more independent and flexible algorithm transplantation, it further enhances the efficiency of software development and avoids duplication of wheel building. It is reported that the above is only the first step of “Cortex” NPU software open source plan, and An Mou Technology will gradually open more resources, such as model quantization, operator implementation and other source code.
The first batch of partners have already joined the Cortex NPU software open source program, some of them are from AIoT, smart cars, intelligent operating systems and other fields.