When AMD Xiaolong meets Alveo, let the CPU do what it is better at

Latest update time：2021-09-03 15:57

Reads：

In the digital age, the development and application of IoT and artificial intelligence technologies have led to IoT devices deployed in homes, cars, airplanes, manufacturing and other links generating massive amounts of data every day, and the data continues to grow at a rate of more than 30%.

When this data is transmitted back to the data center for processing, it places unprecedented pressure on the network bandwidth of the data center's access network and intranet.

You are still using a standard network card

To deploy your server?

If so

The CPU is depressed

…

Standard network cards only have the most basic network processing functions, while complex processing such as network virtualization needs to be completed by the server's CPU. Faced with the above-mentioned surge in network data, the CPU needs to consume a lot of computing resources to perform these network processes. And these computing resources should be used to allocate to real (hard) and (core) applications!!!

The second reason for CPU depression is that the previous generation PCIe Gen 3 standard limits the interaction of data between the network card and the CPU. When the network interface bandwidth is upgraded to 100Gbps or even 200Gbps, the use of PCIe Gen 3 is no longer sufficient.

The third reason for CPU depression is that when facing high-concurrency and high-traffic network data, due to its own architecture, it leads to increased processing delays and slow performance improvement. Increasing the number of CPU cores does not lead to a linear increase in processing performance.

Therefore, in order to make your CPU(s) no longer depressed, the editor needs to recommend to you: AMD EPYC Xiaolong server with Xilinx FPGA smart network card can effectively solve the above bottlenecks.

1. Xilinx's Smart NIC can offload data center network-related processing to the NIC, thereby releasing valuable CPU computing resources for actual applications.

2. Using PCIe Gen 4.0 connection between server CPU and smart network card can double the data transmission bandwidth compared with the previous generation PCIe Gen 3.

3. Hardware offloading has lower latency and higher performance compared to pure software processing.

Recently, AMD and Xilinx and other ecological partners jointly held the "High-speed Interconnection Innovation Unlimited - AMD PCIe 4 Ecosystem & Solution Network Summit". Liu Wenzhuo, Technical Support Director of AMD China Commercial Business Unit, and Sun Zhiyi, Data Center Architect of Xilinx, respectively introduced the principles and solutions of high-speed interconnection from the CPU side and the smart network card side in detail.

Click here to watch the full replay. For more technical videos, please follow Xilinx B Station.