University:
Boston University
Course:
GRS CS 910 | Directed Study: Systems
Academic year:

2021
Views:

274

Pages:

18
Author:

mneelsqia5

Beowulf Cluster! (Intel Micros)! MPI)! Historical Perspective:! First “commercial” QCD machine" First University installation in 1989*! *(from left) Roscoe Giles, Glenn Bresnahan and Claudio Rebbi with CM-2! In late 1980's Thinking Machines Corporation the 64K 1 bit processer CM-2 with performance in excess of 2500 MIPs, and floating point above 2.5Gflops! Killed by Beowulf-clusters! !"#$%&'(&) JLab FNAL BG/L QCD Perfect scaling!! Wilson Dirac Inverter BUT THIS IS NOT SUFFICIENT K. Wilson (1989 Capri) ! “lattice gauge theory could also require a 108 increase in computer power AND spectacular algorithmic advances before useful interactions with experiment ...”" • ab initio Chemistry" 1. 1930+50 = 1980" 2. 0.1 flops ! 10 Mflops" 3. Gaussian Basis functions" vs • ab initio QCD" 1. 1980 + 50 = 2030?*" 2. 10 Mflops ! 1000 Tflops" 3. Clever Collective Variable?" *Hopefully sooner but need $1/Mflops ! $1/Gflops! Disruptive many-core Architectures! • 1/4 CM-2! Nvidia FERMI chip! ! 16 K bit serial PE. ! 512 x 32 bit PE = 16 K bits! • Disruptive QCD Technology! Graphic Processor Units ! Nvidia’s Fermi GPU:" 512 cores x 32 bit FPU! 3-6 GigaByte Mem, 1+ Tflop peak •GPU/BU code sent to Jefferson Lab for 3 million $ ARRA Cluster ! 5 x performance @ 20% extra cost !"#$%&'%()*%!+**'%,--% !"#$%&'()%*+,-.%//0123#4%1(5#66( Hardware (e.g., NVIDIA) \�.�•1 Transistors CUDA Cores DouЫe Precision Floating Point Capabllitv Single Precision Floating Point Capabllity Special Function Units (SFUs) / SM Warp schedulers (Der SM} Shared Memory (per SM) - • r��n'l,1 681 million 128 None 1.4 Ьillion 240 30 FMA ops / clock 3.0 Ьillion 512 256 FMA ops /clock 128 MAD ops/clock 2 240 MAD ops / clock 2 512 FMA ops /clock 1 2 ConfiguraЬle 48КВ or 16 ConfiguraЫe 16 КВ or 48КВ 768 Yes Up to 16 64-Ьit 1 16 16 L 1 Cache (per SM) None None L2 Cache ЕСС Memory Support Concurrent Kernels Load/Store Address Width None No No 32-Ьit None No No 32-Ьit кв кв 4 кв кв CUDA CENTER OF EXCELLENCE

The Problem Of Programming