Huawei Launches the First Full-Stack All-Scenario AI Chip and Solution Ascend 310, Enabling the Smart Future
[October 10, 2018, Shanghai] Huawei has launched the first full-stack all-scenario AI chip Ascend 310 on HUAWEI CONNECT 2018. The conference's theme this year is "Activate Intelligence". As a key part of Huawei's full-stack AI solution, Ascend 310 provides critical support for Huawei's comprehensive AI strategy. Ascend 310 breaks through design constraints such as the power consumption and computing capability. With the greatly improved energy efficiency ratio, Ascend 310 extends AI from the data center to the edge and device, and provides brand-new solutions for application scenarios such as safe city, self-driving, cloud service and IT intelligence, smart manufacturing, and robots, enabling the smart future.
Ascend 310 is a highly efficient, flexible, and programmable AI processor that provides multiple data precisions and supports both training and inference scenarios. Ascend 310 balances AI computing and energy efficiency. In typical configuration, 16 TOPS@INT8 and 8 TOPS@FP16 can be output, while the power consumption is only 8 W. Both high-performance and low-power versions are available, meeting performance and power consumption requirements in different application scenarios. Ascend 310 uses Huawei-developed Da Vinci architecture and extends the application scope of the AI chip. Specifically, this chip provides strong AI computing capability in scenarios where the physical volume and power consumption are limited, greatly supporting all things intelligent. Ascend 310 integrates abundant computing units, which improve the completeness and efficiency of AI computing and therefore expand the application scope of the chip. The programmability of Huawei-developed efficient and flexible instruction sets is capable of embracing both challenges and opportunities brought by the rapid development of AI. With the acceleration of the all-AI service process, the performance of the AI system is greatly improved, while deployment costs are reduced significantly.
With greatly improved computing capability and energy efficiency ratio, Ascend 310 helps extend AI to the edge and device.
Nowadays, AI has entered the high-speed development period. Computing provided by the traditional cloud and data center cannot meet the requirements of mobile devices and other smart terminals. The development of AI to the edge and device becomes inevitable. Based on the previous theoretical basic research on AI, technology accumulation in the field of computing chips, and powerful chip engineering capabilities, Huawei launches the first AI processor Ascend 310 based on the newly self-developed Da Vinci architecture, driving AI extension to the device and edge.
Ascend 310 delivers 16 TOPS@INT8 performance under typical configuration of 8 W power consumption, bringing industry-leading on-device and edge AI computing power. To meet the power consumption requirements in different application scenarios, Ascend 310 provides low-power versions as well.
Newly self-developed Da Vinci architecture greatly improves the computing capability and energy efficiency ratio.
Ascend 310 uses the brand-new Da Vinci computing architecture that is designed for AI computing. Based on the high-performance 3D Cube computing engine, this architecture accelerates matrix computing, greatly improving the AI computing capability in unit power consumption. Each AI core can implement 4096 MAC operations within one clock cycle, an exponential increase in comparison with the traditional CPU and GPU. In addition, to improve the computing completeness and efficiency of AI in different scenarios, Ascend 310 integrates multiple computing units such as the vector, scalar, and hardware accelerator. It also allows computing with multiple hybrid precisions and provides data precisions that meet requirements in training and inference scenarios, achieving AI coverage in all scenarios.
The Da Vinci architecture is flexible and tailorable as well, which can be used in training application scenarios ranging from tens of milliwatts to hundreds of watts. Such flexibility supports all-scenario deployment of AI across the device, edge, and data center, implementing device-cloud homogeneity and enabling device-cloud synergy. Device-cloud synergy means that the training model on the cloud can be seamlessly deployed on the device and edge in real time. Like a human brain, an idea can be executed at the nerve end immediately after being generated. "We get exactly what we think."
The self-developed, efficient, and flexible instruction sets help easily cope with the opportunities and challenges presented by the rapid development of AI.
AI is developing rapidly in both academia and industry, while ASIC chips fail to meet the new requirements posed by this rapid change. To address this issue, Ascend 310 uses efficient and flexible customized instruction sets. Its programmability provides sufficient flexibility to cope with the rapid emergence of new algorithms and models. The instruction sets support forward compatibility, which protects the investment of early algorithm developers. With the flexibility of the instruction sets, Ascend 310 enables AI data scientists to develop new models more quickly and improve AI development to a new level. It also enables AI engineers quickly deploy new services and implement features such as online upgrade of AI, helping achieve new development and breakthroughs in academia and industry.
Computing acceleration and performance upgrade make the ubiquity of AI possible.
To complete an AI task, information needs to be preprocessed and analyzed in addition to the computing of massive data. Ascend 310 is designed from the systematic perspective. Integrated with multiple built-in hardware accelerators, it can accelerate the entire AI service process, including information preprocessing, AI computing processing (AI computing unit), and information post-processing. This greatly improves the performance of the entire AI system and effectively reduces deployment costs, making the ubiquity of AI possible.
The built-in DVPP module processes the input information before it is sent to the AI core computing unit. Therefore, no external processing module is required. The built-in compression and decompression engine can perform lossless compression on the weight and feature map, greatly improving the memory bandwidth utilization and the performance of memory-sensitive applications. The encryption and decryption engine is also integrated to protect data encryption and decryption of the model.
Rich I/O interfaces and multiple chip combinations easily cope with computing challenges from the data center.
One-stop AI integration development suite helps developers enable a smart future.
In addition to Ascend 310, Huawei also provides the one-stop AI integration development suite Mind Studio that applies to the entire AI development lifecycle, from development, optimization, deployment, to maintenance. This lowers the threshold for AI development, reduces AI enabling costs, accelerate the implementation of AI in each industry, helping achieve the vision of all things intelligent.
Enabling AI applications in multiple scenarios to start a smart life.
With the continuous breakthroughs on AI technologies, we are about to enter the era of all things intelligent, where smart life, smart city, and smart transportation will become realities. Based on the improved AI computing capabilities at the device and edge, Ascend 310 will be widely used in various scenarios, such as smart city, self-driving, smart new retail, robot, industrial manufacturing, and cloud computing AI services, and continuously enables AI applications at the device to fully enable the smart life.
First Step of Building a Smart City.
In the application scenario of the smart city, the "eyes" and "brain" of the city need to be upgrade. With many industry chain partners, Ascend 310 will leverage its leading AI computing to transform "eyes" of the city from seeing clearly to seeing through. This is the key to realizing the smart city. Only by building a more secure living environment can we make a smart city a reality.
New Exploration of Self-Driving.
Self-driving is always a hot application, or even "the Everest" in the AI field, attracting the attentions and investment from many industry chain partners. Ascend 310 provides the leading AI computing capability at the edge, which has been fully recognized by vendors at home and abroad. In the future, we will carry out further technology explorations of self-driving to provide more choices for partners.
Fostering Smart Deployment of Cloud Services and IT.
The cloud is the main output platform of the AI computing capability and the basis for implementing the ubiquity of AI. To provide developers with affordable, sound, and assured AI, Huawei starts from the chip design and builds an AI cloud service platform based on different hardware platforms of Ascend 310. Combined with the generality of the AI field, Ascend 310 accelerates the implementation of AI applications such as voice, video, image, and recommendation, greatly reducing AI costs and improving security. Based on the flexibility of the Da Vinci architecture, Ascend 310 can be deployed at the data center, end, and edge, implementing device-cloud homogeneity and device-cloud synergy. This opens up multiple possibilities for new service deployment and implementation.
Enablement of Emerging Industries.
During the implementation of the smart future, industries such as smart manufacturing, smart new retail, and robot have emerged, which present new requirements for AI computing. Based on the precise positioning of Ascend 310 at the device and edge, this chip can effectively enable and promote the development of emerging industries. For data center services, Ascend 310 can be interconnected with multiple chips to constitute different forms of AI acceleration cards, meeting diversified requirements of partners and customers. In addition, the built-in powerful video processing unit reduces the complexity and cost of service deployment, which is recognized and favored by vendors at home and abroad.
Ascend 310 is Huawei's first AI processor based on the new computing architecture Da Vinci. By providing the full-stack AI solution, Ascend 310 can extend AI from the data center to the device and edge and accelerate the implementation of AI in each industry, realizing the vision of all things intelligent.