Adaptive computing accelerates the arrival of the software-defined hardware era

Original Cheng Wenzhi Electronic Enthusiast Network Yesterday
In the past, when designing a product, you must first plan the hardware architecture. After the hardware design is completed, the development of the software part will begin, and then the complete product will be released. Now, with the development of cloud computing, the Internet, and the rise of AI, 5G and autonomous driving, the development process requirements of hardware and products are undergoing unprecedented changes, such as higher hardware performance; higher security and confidentiality requirements; Increasing sensor types and interfaces; constantly evolving AI algorithms and models; and software development needs to be synchronized with hardware development, and so on.

Driven by these new requirements, the concept of "software-defined hardware" has been mentioned many times. People hope that the control and scheduling of all operations in the chip will be completed by software, so as to reduce the corresponding hardware overhead and use the saved part. For computing and on-chip storage. This wish looks very good, but there are still many difficulties to realize it. For example, FPGA can realize some software-defined hardware functions, but its efficiency is lower than ASIC, but its power consumption is higher than ASIC. Is there any better way?
Advantages of adaptive platforms

Xilinx's adaptive computing platform was born for this. According to Xilinx's adaptive computing white paper and adaptive computing area, adaptive computing is based on FPGA technology and supports the dynamic construction of domain-specific architecture (DSA) on the chip. In other words, adaptive calculation allows DSA to be dynamically updated with changes in demand, thereby avoiding the constraints of long ASIC design cycles and high NRE costs. With the continuous improvement of the distributed level of processing, adaptive computing can not only support the over-the-air (OTA) update of software, but also the over-the-air update of hardware, and the update can be repeated almost wirelessly.
"Adaptive platform" refers to any type of product or solution centered on adaptive hardware. The adaptive platform is completely based on the same adaptive hardware foundation, but it contains far more than chip hardware or devices, but covers all hardware and a comprehensive set of design software and operating software.

"Adaptive Computing White Paper"

Scan code to download now
With an adaptive platform, hardware engineers can free themselves from repetitive and low-end design work and focus on the development of professional functions that they are good at. Software engineers can start design work at the same time as hardware engineers without waiting. After all the hardware is designed, it starts to work.

Figure: Schematic diagram of unconfigured and configured adaptive hardware (source: Xilinx)
Of course, in addition to this benefit, the adaptive platform has the following advantages:
One is to speed up the product launch process. For example, one of Xilinx's adaptive computing platform products, the Alveo data center accelerator card, uses the accelerator card to build applications that can be accelerated for specific applications without the need for special hardware customization. And just add the PCIe card to the service, you can directly call the acceleration library from the existing software application.
The second is to reduce operating costs. Compared with CPU-based solutions, due to the increase in computing density, optimized applications based on adaptive platforms can provide significantly improved efficiency at each node.
The third is that the workload can be configured flexibly and dynamically. The adaptive platform can be reconfigured according to current needs. Developers can easily switch deployed applications within the adaptive platform, and use the same device to meet changing workload requirements.
The fourth is to be compatible with the future. The adaptive platform can continuously adjust. If existing applications require new functions, the hardware can be reprogrammed to implement these functions in the best way, reducing the need for hardware upgrades, and extending the service life of the system.
Fifth, the overall application can be accelerated. Because AI inference rarely exists alone, it is generally part of a larger data analysis and processing chain, often coexisting with multiple upstream and downstream stages that use traditional (non-AI) implementation solutions. Embedded AI in these systems partly benefits from AI acceleration, while non-AI parts can also benefit from acceleration. The natural flexibility of adaptive computing is suitable for accelerating AI and non-AI processing tasks, which is called "overall application acceleration". As computationally intensive AI inference penetrates into more applications, its importance is also increasing.
Successful landing case of adaptive computing

In the past, if engineers wanted to use FPGAs, they needed to build their own hardware boards and configure FGPA with a hardware description language (HDL). Nowadays, developers of adaptive platforms only need to use their familiar software frameworks and languages ​​(such as C++, Python, TensorFlow, etc.) to directly exert the effectiveness of adaptive computing. In other words, software and AI developers do not need to build circuit boards or become hardware experts to use adaptive computing freely.

What's more convenient is that engineers can not only directly call their existing software code through APIs, but also use the open source libraries provided by the independent software vendor (ISV) ecosystem and vendors. There are a large number of accelerated APIs available in the library.
Take the two adaptive computing platform products Kria SOM and Alveo accelerator cards that Xilinx has mass-produced as examples. Kria SOM is built on the Zynq UltraScale+ MPSoC architecture and supports developers to develop edge applications on a turnkey adaptive platform. By standardizing the core parts of the system, developers have more time to focus on creating differentiated features.
 

Xilinx’s first mass-produced Kria SOM product is the K26 SOM. In terms of hardware configuration, the K26 SOM is based on the Zynq UltraScale+ MPSoC architecture design, with an overall size of 77×60×11mm, equipped with a quad-core Arm A53 processor, and a built-in 64-bit 4GB DDR4 memory, with 256K system logic unit and 1.4TOPS AI processor performance, supports 4K 60p H.264/265 video codec.
Kria SOM is designed, manufactured and tested as a mass production-ready product, which can withstand a variety of harsh application environments. At present, Kria SOM is divided into two categories: industrial grade and commercial grade. The industrial grade supports higher vibration and more extreme temperatures, as well as longer life cycle grades and maintenance.
 

Kria SOM is mainly for intelligent vision applications. Therefore, it can be used for high-speed target detection in smart cities, such as license plate recognition. At the same time, it can also be used for machine vision applications on industrial production lines.
For the Alveo accelerator card, it uses the industry-standard PCI-e interface, which can provide hardware offloading capabilities for any data center application, and can also be used for SmartSSD storage to accelerate on storage access points. In addition, it can also be used for SmartNIC to directly provide acceleration on network traffic.
For example, Alveo SN1000 SmartNIC, which expands the performance envelope of SmartNIC, targets data centers and edge computing platforms, combines high-performance networks, CPU clusters and large-scale FPGAs, and builds a high-performance computing (HPC ) The platform has a significant network acceleration function.
In addition, Alveo SN1000 SmartNIC adopts standardization and software framework, does not need to deal with FPGA programming directly, it is more convenient to use. Engineers can use Xilinx or third parties to support most of the firmware used in FPGAs, and even software running on CPU clusters. The CPU cluster can run standard Linux distributions, such as Ubuntu and Yocto Linux. The SmartNIC driver can be used on host platforms such as Red Hat Enterprise Linux (RHEL), CentOS and Ubuntu.
In terms of application, Alveo is suitable for genomics analysis, graphics database, medical image processing and analysis, and video-based image monitoring applications. In terms of application landing, there are already applications in data centers and gene sequencing applications.