A New Golden Age for FPGA in Data Centers

Publisher:闻天明Release Time:2020-05-22Number of visits:277

You can also attend this talk online; the talk will be conducted in Chinese. Tencent Meeting : 519 056 233
 -----------------------------
 Long time no see: )
 -----------------------------
 Speaker:  Dr. Bojie Li

Time:       15:00-16:30, May. 27/ Tencent Meeting : 519 056 233

Location:  SIST 1D 106

Host:       Prof. Yajun Ha

 

Abstract:

Data centers are the infrastructure that hosts Internet services all around the world. But now they are facing challenges on hardware and application.

On the hardware side, performance improvement of general processors is slowing down. On the application side, big data and machine learning require increasingly high computational power, while the workload patterns are significantly different.

Achieving significant performance gains require innovations both hardware architecture and hardware-software interface. As a result, domain specific architectures (DSAs), such as GPU, TPU, and FPGA, attract much interest in data centers. Different DSAs follow diverse architectural philosophies and optimize for distinct types of applications.

This talk will focus on FPGA, an interesting DSA that are deployed by many cloud data centers, including Microsoft Azure, Baidu, Amazon AWS, Tencent, Alibaba and Huawei. FPGA is interesting because it allows the user to reconfigure the hardware architecture on the fly, so it is very flexible. The flexibility of FPGA makes its way into two major application scenarios: cloud infrastructure acceleration and tenant application acceleration.

In the first scenario, FPGA accelerates the services and virtualization abstractions that power a public cloud. Network virtualization, network function virtualization, remote storage access, and key-value stores are several notable examples. In this scenario, FPGAs are hidden from the tenants, and mostly used as a replacement for specialized chips (ASICs).

In the second scenario, the tenants of a public cloud use the FPGAs explicitly by developing custom hardware logic that runs on the FPGA. There are some important use cases: deep neural networks, search ranking, encryption, and image processing. In this scenario, FPGAs look like other accelerators such as GPUs, but differ in suitable applications, interconnect architecture, and programming interfaces.

In addition to the typical applications of FPGAs in data centers, this talk will cover programming FPGAs with high-level languages, and discuss the opportunity for software ecosystem of FPGAs in data centers.

Bio:

Bojie Li (NgZSpg) is a senior developer in Distributed and Parallel Software Lab, Central Software Institute, Huawei 2012 Labs.

In 2019, Bojie Li received his Ph.D. in Computer Science from University of Science and Technology of China (USTC) and Microsoft Research Asia (MSRA), supervised by Prof. Lintao Zhang and Prof. Enhong Chen.

Bojie's research has been advancing the state of art in high performance data center systems with reconfigurable hardware. During Ph.D. study, Bojie was an intern at Microsoft Research Asia, and played a significant role in the research of SmartNIC applications in data centers.

Bojie has published papers on top networking and systems conferences, including SIGCOMM, SOSP, NSDI, and ATC. Besides, he has received ACM China Doctoral Dissertation Award, Chinese Academy of Science Presidential Scholarship, and Microsoft Research Asia Fellowship Award

SIST-Seminar 18237