How to deploy a deep-dimensional FPGA+CPU image processing solution in a data center

Editor’s note: A technological breakthrough has been achieved, so how to integrate with the production environment? How to better apply the advantages of FPGA in the actual business system? The actual deployment is the most worthy of discussion and research direction. In this regard, Shenwei Technology has carried out a series of explorations and attempts, and formed the following solutions.

1. Deployment issues

1.1 How to integrate with production environment

How to deploy a deep-dimensional FPGA+CPU image processing solution in a data center

In the cloud, ThunderImage is usually integrated with OBS, the client sends a request for the image, the CDN checks whether there is a local cache, if not, it sends a request to OBS, and OBS interacts with ThunderImage through corresponding parameters to generate the corresponding image format, and finally Return to the client through CDN. This online processing method integrated with OBS brings a reduction in bandwidth costs and storage costs, as well as an improvement in QoS.

1.2 FPGA product deployment

How to deploy a deep-dimensional FPGA+CPU image processing solution in a data center

At present, FPGA products can be divided into public cloud solutions and private cloud solutions in terms of deployment. The product forms of the solutions currently owned by Shenwei Technology are mainly divided into public cloud: SaaS, Instance and mirroring services; private cloud and offline products: software and hardware all-in-one. The applied scenarios include mobile phones, e-commerce, social networking, CDN, cloud storage, video, etc.

1.3 Private Cloud Deployment Solution

How to deploy a deep-dimensional FPGA+CPU image processing solution in a data center

For private cloud deployment, the architecture is based on the hardware environment of the video accelerator card, and the corresponding operating system and virtualization processing on the upper layer of the server are combined, covering the Xilinx driver layer, management tools and other parts. Shenwei Technology’s solution is deployed on this architecture, directly facing the corresponding application scenarios.

1.4 Alveo U50 Deployment Solution

How to deploy a deep-dimensional FPGA+CPU image processing solution in a data center

Alveo U50 adopts UltraScale+ architecture and integrates ultra-high bandwidth 8GB HBM2 memory technology, so the product size is greatly reduced. The solution based on this is the current processing solution with the highest computing density, which is more convenient for data center server upgrades.

2. Case study

How to deploy a deep-dimensional FPGA+CPU image processing solution in a data center

The above picture is an actual case of a social network cloud album thumbnail production. The throughput performance of a single node can be improved by 16 times, the TCO can be reduced to half, and it has a better service quality experience. Because its cluster size is smaller, it is easier to maintain.

2.1 Public cloud deployment

How to deploy a deep-dimensional FPGA+CPU image processing solution in a data center

Public cloud deployment mainly involves platforms, functions, and services. Among them, the service forms on the cloud include SaaS, Instance, and Image. The advantage of SaaS is that it is easy to deploy, has a low start-up cost, and is convenient for small-scale users to integrate and use. The integration of Instance is more flexible, the package is more complete, and the cost is lower than that of SaaS. For Image, the purchase of cloud hosts and images requires the perfect integration of applications and images on the server side, which means that we can have the opportunity to bring out the best performance of the overall solution. Therefore, the performance of the Image method is the strongest, and Deep customization is possible.

2.2 WebP transcoding scheme of a video website

The above picture shows the actual case of a video website’s WebP transcoding solution. Our overall performance on the cloud is guaranteed. Although the CPU in the cloud is a VCPU, there will be some performance losses, but the overall performance gains will generally be More than 20 times the performance improvement.

To sum up, the ThunderImage image processing solution of Shenwei Technology supports a wealth of encoding and decoding algorithms, can adapt to differentiated application scenarios, and is seamlessly compatible with common algorithm platforms in the industry. Efficient and convenient application experience and deployment methods are a new optimization and choice for applications in public cloud, private cloud or hybrid cloud.

About Shenwei Technology:

Beijing Shenwei Technology Co., Ltd. (Shenwei Technology for short) was established in 2016 and is composed of top FPGA software and hardware developers in China. The company’s team has extensive experience in multimedia processing, HPC applications and heterogeneous system architecture design. Committed to growing into a leading FPGA computing platform supplier for heterogeneous acceleration applications, providing support for more industries and applications.

Based on FPGA + CPU heterogeneous computing technology, Shenwei Technology provides ultra-high-performance image and video processing solutions and products for data center applications. In addition, Shenwei Technology also provides solutions and design services for high-performance computing and big data applications, such as storage compression acceleration and seismic exploration.

The Links:   LTD104C11U LQ065T5BR02 IGBTMODULE

Related Posts