Author: Greg Martin, director of strategic marketing, Xilinx

(Xilinx is now part of AMD)

The constantly changing and evolving 5g, data center, automotive and industrial applications require continuous improvement of computing capacity while maintaining a rigorous power envelope. As the business process of artificial intelligence (AI) technology continues to accelerate, it has become a major factor to improve the computing density.

Whether deployed in the cloud, edge or terminal, artificial intelligence inference requires higher processing performance and strict power budget. Therefore, artificial intelligence inference workload usually needs special artificial intelligence hardware to accelerate.

At the same time, the development speed of artificial intelligence algorithms has far exceeded the speed of the traditional chip development cycle. Due to the rapid innovation of advanced artificial intelligence models, fixed chip solutions, such as the ASIC implementation of artificial intelligence networks, may soon be eliminated.

Adaptive computing is the answer to these challenges

Because it is built on adaptive hardware that can still be optimized for specific applications after product manufacturing, adaptive computing has unique value. Since optimization can be carried out on demand after the completion of hardware manufacturing, it can keep pace with the latest artificial intelligence models. On the contrary, because ASIC is based on fixed hardware architecture, it cannot be changed once manufacturing is completed.

This flexible optimization capability of adaptive computing can support infinite repeated execution. Even after the device is fully deployed to the mass production environment, the hardware can still be changed. Just as a mass production CPU can be used to run a new program, an adaptive platform can also flexibly adapt to new hardware configurations, even in a real-time production environment.

Comparison of adaptive hardware with other alternatives

CPU and GPU have their own unique capabilities, which are very suitable for some tasks. CPU is the best choice for decision-making functions that need to evaluate complex logic. GPU is the best choice for processing offline data with high throughput but low delay requirements. Adaptive computing is the best choice for those applications that require high throughput and low delay data processing at the same time, such as real-time video streaming, 5g communication and automotive sensor fusion.

The reason why adaptive computing can provide high performance with low delay is that it can implement domain specific architecture (DSA), so as to ensure the best implementation of specific applications in specific domain architecture. On the contrary, CPU and GPU are based on fixed and von Neumann architecture, and it is not allowed to optimize their underlying architecture for specific fields.

DSAs can also be built using dedicated (fixed) chip devices, which are often referred to as application specific standard products or ASSPS. However, implementing DSA in a fixed ASSP has both advantages and disadvantages. Here are two main disadvantages.

The first is the pace of innovation. In order to keep up with the pace of innovation, manufacturers are expected to create and provide new services in a shorter time. More specifically, this time is shorter than the time required to design and develop a new fixed chip DSA. This has caused the fundamental market dislocation between the innovation demand of the market and the time required for enterprises to design and manufacture ASSP. Changes in industry standards or other demand fluctuations will quickly lead to the obsolescence of these devices.

The second consideration is the cost of customized chips. The one-time cost of designing and manufacturing unique chip designs (such as complex 7Nm ASIC) may lead to hundreds of millions of dollars of non repetitive Engineering (NRE) costs. As the device process is reduced to 5nm and smaller, the cost is expected to rise further. Rising costs are delaying the adoption of advanced nodes by ASSP, which may lead its users to stick to outdated and inefficient technologies.

Introduction to adaptive computing platform

Adaptive platforms are all based on the same adaptive hardware (FPGA). However, they contain far more components and technologies than the chip hardware and devices themselves. The adaptive platform includes a comprehensive set of runtime software. The combination of software and hardware provides a unique ability to create highly flexible and efficient applications.

The adaptive platform enables adaptive computing to be used by a wide range of software and system developers, and lays a foundation for them to create many innovative products. The advantages of using an adaptive platform include:

  • Shorten time to market. Using alveo ™ Platforms such as data center accelerator cards can support applications built with hardware specially built for specific application acceleration without customizing hardware. Moreover, simply connect the PCIe card to the server, and you can directly invoke the acceleration library with existing software applications.
  • Reduce operating costs. Compared with the CPU based solution, due to the increase of computing density, the optimization application based on the adaptive platform can greatly provide the efficiency of each node.
  • Flexible and dynamically changing workloads. The adaptive platform can be reconfigured according to current requirements. Developers can easily switch deployed applications within the adaptive platform, and use the same device to meet the changing workload requirements.
  • Compatible with the future. The adaptive platform can be adjusted continuously. If the existing application needs new functions, the hardware can be reprogrammed to realize these functions in the best way, reduce the hardware upgrade requirements, and then extend the service life of the system.
  • Accelerate overall application. AI inference rarely exists alone. It is part of a larger data analysis and processing chain and often coexists with multiple upstream and downstream stages using traditional (non AI) implementation schemes. The embedded AI portion of these systems benefits from AI acceleration, while the non AI portion also benefits from acceleration. The natural flexibility of adaptive computing is suitable for accelerating AI and non AI processing tasks, which is called “overall application acceleration”. With the penetration of computing intensive AI inference into more applications, the importance of “overall application acceleration” is also increasing.
  • Ease of use. In the past, using FPGA technology required developers to build their own hardware boards and configure FPGAs with hardware description language (HDL). In contrast, the adaptive platform supports developers to use their familiar software frameworks and languages (such as c+++, python, tensorflow, etc.) to directly exert the efficiency of adaptive computing. Software and AI developers can now use adaptive computing directly without building circuit boards or becoming hardware experts.

Different types of adaptive computing platforms

According to applications and requirements, there are many types of adaptive platforms, including data center accelerator cards and standardized edge modules. The existence of multiple platforms aims to provide the best possible starting point for the development of required applications. Different adaptive platforms face a wide range of applications, including time delay sensitive applications such as autopilot and real-time video streaming, as well as highly complex 5g signal processing and data processing of unstructured databases.

Adaptive computing can be deployed to the cloud, network, edge and even terminal, bringing the latest architecture innovation to individual and end-to-end applications. In view of the existence of various adaptive platforms, the deployment locations can also be diverse – from high-capacity devices on PCIe accelerator cards in the data center to small low-power devices suitable for terminal processing required by IOT devices.

Adaptive platform at the edge, including Xilinx KRIA ™ Adaptive system module (SOM), the adaptive platform in the data center includes the alveo accelerator card. Alveo accelerator card adopts industry standard PCIe, which provides hardware unloading capability for any data center application.

KRIA adaptive som

Introduction of AI engine

One of the biggest innovations in the field of adaptive computing is the AI engine launched by Xilinx.

AI engine is a revolutionary new method, which provides unprecedented computing density for computing intensive applications. The AI engine is still fundamentally a configurable block, but it can also be programmed like a CPU. The AI engine is not composed of standard FPGA processing hardware, but contains high-performance scalar and single instruction multiple data (SIMD) vector processors. These processors are optimized to efficiently implement the various compute intensive functions that occur in artificial intelligence inference and wireless communication.

The AI engine array is still connected to the flexible and flexible data interconnection similar to FPGA, so it can establish an efficient and optimized data path for the target application. This computing intensive, CPU like processing element and FPGA like interconnection combination is leading artificial intelligence and communication products into a new era.

Xilinx AI engine architecture

Welcome a more interconnected and intelligent world

Fundamentally speaking, adaptive computing is based on the existing FPGA technology, but it is easier to be accepted by more developers and applications than ever before. Software and artificial intelligence developers can now quickly build optimized applications with adaptive computing hardware technology, which was once out of reach for them.

The ability to adapt hardware to specific applications is the unique difference between adaptive computing and CPU, GPU and ASSP. The core of the latter is the fixed hardware architecture. Adaptive computing allows the hardware to be tailored to the application to achieve greater efficiency, and it can also be adjusted according to needs if future workloads or standards change.

As the world becomes more interconnected and intelligent, adaptive computing will continue to occupy the forefront of optimizing and accelerating applications, helping all kinds of developers accelerate the transformation of ideas into reality and make our tomorrow better.

Leave a Reply

Your email address will not be published.