With the development of deep learning and neural network technology, FPGA acceleration technology is more and more widely used in the fields of artificial intelligence, 5g, automatic driving and data center. FPGA supports differentiated customization and is a competitive accelerator for big data, deep learning and wireless communication; FPGA supports the re adjustment of underlying hardware architecture and software customization, so it can integrate the latest industry innovative technologies; FPGA has the characteristics of efficient and repeatable programming, which can realize customized performance, customized power consumption, high throughput and low batch delay, and meet the requirements of various specifications of users.
FPGA has the following technical advantages: (1) high performance and low power consumption. FPGA is a programmable chip with high performance and low power consumption. Targeted algorithm design can be made according to customer customization. When dealing with massive data, compared with CPU and GPU, FPGA has the advantage of higher computing efficiency, and its high parallelism can often improve the business performance by an order of magnitude.
It is a device integrating software and hardware without instructions and software. FPGA programming uses hardware description language, and the logic described by hardware description language can be directly compiled into the combination of transistor circuits. Therefore, FPGA actually uses transistor circuit to realize user algorithm directly, without the translation of instruction system.
(2) It has strong flexibility. FPGA is a field programmable logic gate array and a combination of a pile of logic gates. It can be programmed or reprogrammed. FPGA is dynamically reconfigurable. After the data center is deployed, different logic can be configured according to the business form to realize different hardware acceleration functions. For example, the FPGA board on the current server deploys picture compression logic to serve social instant messaging; At this time, advertising real-time prediction needs to expand capacity to obtain more FPGA computing resources. Through a simple FPGA reconfiguration process, FPGA board can be transformed into “new” hardware to serve advertising real-time prediction, which is very suitable for batch deployment.
In the future, FPGA will become an increasingly important chip. Heterogeneous computing with deep integration of CPU and FPGA has become an inevitable choice for industrial development. The heterogeneous multi-core computing innovation of CPU + FPGA involves deep-seated integration innovation such as software and hardware, which provides a new runway for catching up and surpassing. At the same time, the heterogeneous computing of CPU + FPGA, as a new important technology trend, not only accelerates the integration of computing architecture, but also brings important opportunities to the downstream machine system and software development. In the era of big data, machine learning and deep learning are the main driving forces of artificial intelligence. In recent years, the rapid growth of big data and computing power has made a qualitative leap in deep learning technology, resulting in major breakthroughs in the fields of computer vision, speech recognition, natural language processing and so on.
However, deep learning requires a lot of parallel computing and has high requirements for hardware platform, which can not be met by traditional computers. FPGA has high performance power consumption ratio, and based on gate circuit design, FPGA is a scheme with ultra-low delay and definite delay. The programmability and dynamic reconfigurability of FPGA can adapt to the changes of future algorithms for deep learning, and IO programmability can meet more business requirements (network acceleration, edge computing), Therefore, the deep learning scheme based on FPGA has become the development direction of technology in the future.
As a powerful computing accelerator in the future, FPGA will not only affect the decision-making and market trend of major enterprises, but also accelerate the workload in enterprises, promote the internal search of large-scale data centers, and improve the status of HPC high-performance computing simulation.
A new revolution represented by the technologies of Internet of things, intelligent manufacturing, big data, cloud computing and artificial intelligence is in the ascendant. In this context, the Chinese government has put forward the development strategies of “intelligent manufacturing” and “scientific and technological innovation in the 13th five year plan” to vigorously promote industrial transformation, upgrading and structural adjustment. Facing the R & D and application environment with rapid technology update and frequent iteration, FPGA (field programmable gate array) is becoming an increasingly important chip and has been more and more widely used. It plays an important role in the fields of Pan artificial intelligence, 5g, unmanned driving, intelligent terminal and data center.
In an interview, the CEO of an AI chip startup company said, for example, in deep learning, it is often necessary to adjust the algorithm. If ASIC is adopted, it is possible that the traditional algorithm hardening has just been completed, and after half a year, the algorithm team changes the algorithm again. The flexibility of using FPGA can well adapt to this change. Audi is using Intel FPGA to achieve automatic driving of Audi A8 car, and Denso has already adopted Intel technology driven stereoscopic visual effect system. More and more automobile companies use FPGA in their vehicles to provide functions such as adaptive cruise control, collision avoidance and driver assistance.
FPGA is also widely used in large cloud data centers to realize artificial intelligence applications. For example, Microsoft’s brainwave project is an accelerated deep learning platform based on FPGA. It uses Intel Stratix FPGA to provide real-time artificial intelligence in the cloud through data processing and transmission as soon as possible. Dell EMC is using Intel FPGA chips for application verification in the fields of gene analysis and protein structure research.
PowerEdge r940xa server is designed to accelerate business critical application databases without the cost and security risk of using the cloud. It combines four CPUs with four graphics processing units (GPUs) at a large ratio of 1:1, improves application performance, and realizes low latency through directly connected nvme drivers. The PowerEdge r840 server is designed for in database analysis. It uses more direct nvme drives than other server products on the market to minimize data latency and accelerate data transmission through a fully integrated super channel interconnect bus (UPI).
Dell Eason also participated in the Intel AI builder program, an ecosystem for enterprise technology partners to develop high-quality solutions based on Intel AI products. Dell Eason cooperates with Intel to help customers evaluate their position in Intel through Dell Eason innovation laboratory and Dell Eason customer solution center ® xeon ® Scalable processors and Intel ® Large scale deep learning and high-performance computing workload on FPGA.