臺灣博碩士論文加值系統

摘 要
適用於半導體業界多年的摩爾定律已使得今日廣被使用於雲端資料中心伺服器裡的通用型處理器(General Purpose CPU，如 x86 CPU)的晶片整合度，如更多的處理器的內核數( internal processor core)，內部快取記憶體及內建記憶體控制器通道(Memory Channel)等， 達到前所未見的程度。
然而，在使用者端看到的只是CPU功耗的增加及更多的散熱需求外，受益到的效能提升卻是有限。當通用型CPU的高度元件整合，功率消耗卻無法明顯改善的情況下， 而半導體往先進製程推演進速度間隔變得更久，到了驅動半導體界發展的摩爾定律所預測的積體電路上可容納的電晶體數目，約每隔兩年便會增加一倍每十八個月只至兩年的時間間隔，變得難再適用。
有鑑於處理器效能與功耗的改善緩慢，近年來硬體加速與CPU協同運作的異構運算架構, 用於雲端資料中心等伺服器裡，日漸興起。搭配專屬加速功能的硬體加速器用來加速特定應用，以期能增進總體雲端伺服器的效能與改善功耗，藉由可平行化的跨平台程式編譯，將特定工作offload 給CPU以外的協同處理器來做，例如CPU與GPＵ的異構協同運算，GPU專注在圖形運算上，或高精度浮點運算上而CPU可以集中處理作業系統或網路資料的傳送。再如CPU 與FPGA (Field Programmable Gate Array) 的搭配，FPGA專門用來進行CNN (Convolution Neural Network) 類神經網路的連乘加運算，以加速DL (Deep Learning) 深度學習的速度。
本論文旨在對GPU與FPGA架構上的探討，歸納FPGA在效能上能夠與GPU相比擬，加上FPGA廠商提供開發環境的完備及開源的跨平台OpenCL程式語言，以及可以重新組態的特性， 針對特定應用而客製化，傳統雲端硬體供應商將會有機會來開發不同加速應用的FPGA卡用以差異化，增強產品競爭力。

Abstract
Major server CPU (Central Processing Processor) processor nowadays are capable of equipping more internal processor cores, greater than ever internal cache and memory controller channels. However, with such highly integration of the micro-processor evolution, the users see only limited performance gains but much more burden of increasing processor TDP (Thermal Design Power), heat dissipation and cooling requirements. It is observed that evolution course of semiconductor process has reached to where Moore's Law called out decades ago that transistor counts double every 18 to 24 months can no longer be applied.
Recent years in data center industry, hardware accelerators with specialized processing unit have been proposed and deployed in some cloud computing architectures to accelerate specific applications for increasing performance and improving power consumption. As a result, heterogeneous computing in datacenter has aroused, such as collaboration of CPUs and GPU while GPU focus on graphics operations or high-precision floating-point operations. The CPU focus on proceeding other operation system works or network data movement.
This study is aimed to understand GPU and FPGA fundamental architecture difference and their performance dominance, how FPGA can be comparable to GPU from the view point of cloud-based applications performance. In addition, FPGA’s reconfigurable characteristic provides a great deal of design flexibility to be suitable for diverse applications acceleration need in the cloud and an alternative solution to GPU for acceleration. Given the thorough development environment FPGA vendors provide and open source parallel programming language like OpenCL, hence server hardware providers will be able to offer FPGA acceleration to fulfill their diverse cloud customers need, more importantly to differentiate product portfolios and enhance the competitiveness.

Contents
摘 要 I
Abstract II
Contents III
Tables IIII
Figures V
Chapter 1 Introduction 1
1.1 Background 1
1.2 Moore’s Law obsolete 2
1.3 Research Methodology 4
Chapter 2 Heterogeneous Computing 5
2.1 Industry approach 5
2.2 Performance and power consumption consideration 6
2.3 GPU 7
2.3.1 Inside GPU 9
2.3.2 GPU Memory Management 10
2.4 FPGA 11
Chapter 3 Cloud Acceleration 19
3.1 Cloud Computing Era 19
3.2 Cloud Applications characterization 20
3.3 Reconfigurable Acceleration for applications 21
Chapter 4 Market Overview 24
4.1 Brief History of Servers 24
4.2 Market Momentum 25
4.3 Competition 26
4.3.1 Opportunity 27
4.3.2 Brand vendors opportunity 29
Chapter 5 Near Future Observation 33
Chapter 6 Conclusion…………………………………………………………37
References 36

References
[1] Andrei-Alexandru Dafinoiu, “INVESTIGATION OF HETEROGENEOUS COMPUTING THROUGH NOVEL PARALLEL PROGRAMMING PLATFORMS, 2016
[2] Power consumption based on models in “Dark Silicon and the End of Multicore Scaling,” Hadi Esmaelizadeh, ISCA, 2011
[3] Hadi Esmaeilzadeh et al. “Dark Silicon and the End of Multicore Scaling”, 2011
[4] Onur et al. µC-States: Fine-grained GPU Datapath Power Management, 2016
[5] Babak Falsafi, Bill Dally, Desh Singh, Derek Chiou, Joshua J. Yi, Resit Sendag, "FPGAs versus GPUs in Datacenters," IEEE Micro, vol. 37, no. 1, pp. 60-72, Jan 2017
[6] Cathal Murphy et al.”Xilinx All Programmable Devices :A Superior Platform for Compute-Intensive Systems”,2017
[7] Fowers, J., Ovtcharov, K., Strauss, K., Chung, E.S., Stitt, G.: A High Memory Bandwidth FPGA Accelerator for Sparse Matrix-Vector Multiplication. In: IEEE Int. Symp. on Field-Programmable Custom Computing Machines (2014).
[8] Yaman Umuroglu, et al. "Random Access Schemes for Efficient FPGA SpMV Acceleration." Microprocessors & Microsystems Volume 47 Issue PB, November 2016 Pages 321-332
[9] NVIDIA Corporation. "Cuda C Programming Guide" Last accessed on April 6, 2017.
[10] William Dally. "High-Performance Hardware for Machine Learning." Last accessed April 6, 2017.
[11] Mark Gebhart et al. "Unifying Primary Cache, Scratch, and Register File Memories in a Throughput Processor." MICRO-45 Proceedings of the 2012 45th Annual IEEE/ACM International Symposium on Microarchitecture, Pages 96-106, Vancouver, B.C., CANADA - December 01–05, 2012.
[12] Chen, Doris, and Deshanand Singh. "Invited paper: Using OpenCL to evaluate the efficiency of CPUS, GPUS and FPGAS for information filtering."Field Programmable Logic and Applications (FPL), 2012 22nd International Conference on. IEEE, 2012.
[13] Xilinx white paper, "Xilinx Reduces Risk and Increases Efficiency for IEC61508 and ISO26262 Certified Safety Applications." Last accessed April 6, 2017
[14] ANANDTECH. "NVIDIA Announces Tesla P40 & Tesla P4 - Neural Network Inference." Last accessed April 6, 2017.
[15] Xilinx white paper, ''Deep Learning with INT8 Optimization on Xilinx Devices." Last accessed April 6, 2017
[16] Xilinx, Inc. "Partial Reconfiguration in the Vivado Design Suite." Last accessed April 6, 2017.
[17] Song Han et al. "ESE: Efficient Speech Recognition Engine with Sparse LSTM on FPGA." International Symposium on FPGA 2017.
[18] Jian Ouyang et al. "SDA: Software-Defined Accelerator for General-Purpose Big Data Analysis System." Hotchip 2014. 

[19] Shreyas G Singapura et al. "FPGA Based Accelerator for Pattern Matching in YARA Framework." CENG 2015
[20] Yuichiro Utan et al. "A GPGPU Implementation of Approximate String Matching with Regular Expression Operators and Comparison with Its FPGA Implementation." PDPTA 2012.
[21] Microsoft research, "A Cloud-Scale Acceleration Architecture." Last accessed April 6, 2017.
[22] S. Che et al., “Accelerating compute-intensive applications with gpus and fpgas,” in SASP’2008, pp. 101–107
[23] K. Krommydas et al., “On the characterization of opencl dwarfs on fixed and reconfigurable platforms,” in ASAP’2014, pp. 153–160.
[24] H. R. Zohouri et al., “Evaluating and optimizing opencl kernels for high performance computing with fpgas,” in SC’2016, pp. 35:1–35:12.
[25] Jason Cong et al. Understanding performance Differences of FPGAs and GPUs 2018
[26] “Cisco Global Cloud Index : Forecast Methodology” 2015
[27] E. Savitz, “Bridging The Data Deluge Gap,” Forbes Tech, August 2012. 

[28] M. Ferdman, A. Adileh, O. Kocberber, S. Volos, M. Alisafaee, D. Jevd- jic, C. Kaynak, A. D. Popescu, A. Ailamaki, and B. Falsafi, “A case for specialized processors for scale-out workloads,” IEEE Micro’s Top Picks, 2014
[29] “The Xilinx SDAccel Development Environment: Bringing The Best Performance/Watt to the Data Center,” Tech. Rep., 2015. 

[30] Christoforos Kachris et al. “A Survey on Reconfigurable Accelerators for Cloud Computing”, 2016
[31] A. Putnam et al.” A Reconfigurable Fabric For Accelerating Large-Scale Datacenter Services”, IEEE Press 2014
[32] Digitimes 2017~2022 global servers forecast growth rate 2017 Q4
[33] “Global Server Industry Market Development Analysis” – MIC 2017 Q4
[34] IDC WW Quarterly Server Tracker 2018
[35] https://www.nextplatform.com/2018/04/11/dell-emc-and-fujitsu-roll-intel-fpgas-into-servers
[36] Digitimes 2018 Taipei Computex Report
[37] Maria Avgerinou et al. Trends in Data Centre Energy Consumption under thewww.hpe.com
[38] Cisco Global Cloud Index, 2016-2021
[39] www.dell.com
[40] www.hpe.com
[41] Intel Acceleration Stack for Intel® Xeon® CPU with FPGAs Workshop