Summary 1
Catalyst 1
Ovum view 1
Key messages 2
The GPU has expanded from graphics processing to number crunching 2
Non-graphic uses for GPUs are already well established in other sectors 2
Many leading capital markets companies are already considering GPUs for their HPC environments 2
GPU performance continues to improve 2
Some software vendors and service providers ease GPU adoption 2
Table of Contents 3
The GPU Has Expanded From Graphics Processing to Number Crunching 4
What is a GPU? 4
What is GPGPU? 4
What are the challenges in GPGPU? 5
GPGPU is Already Well Established in Other Sectors 5
Many Leading Capital Markets Companies are Already Considering GPGPU 6
Finite difference versus Monte Carlo simulation 7
Single versus double precision 7
GPU Performance Continues to Improve 8
Some Software Vendors and Service Providers Ease GPU Adoption 13
Numerical Algorithms Group 13
SciComp 13
Hanweck 14
Bloomberg 14
End User Case Study: BNP Paribas 15
The Competitive Landscape 15
Nvidia is the clear leader in enabling GPGPU 15
Nvidia has also struck OEM partnerships to ease penetration of the HPC market 16
OpenCL will change the competitive landscape during 2012 16
Nvidia's competitors are currently lagging behind in their GPGPU efforts 17
IBM has cancelled the Cell project 17
Intel has dropped Larrabee in favor of MIC chips 17
For the time being, AMD's graphics priorities are elsewhere 18
Trends in GPGPU HPC Applications 19
Recommendations 21
Recommendations for enterprises 21
Define your philosophical stance regarding early adoption 21
Some companies equate early adopter status with first mover advantage, while others see fools rushing in where angels fear to tread. Capital market participants considering GPGPU should ask themselves a couple of questions: 21
Do you consider GPUs as too esoteric?• 21
Are you concerned with proprietary technology tie-in?• 21
Assess the cost implications of migrating applications to GPUs 21
Consider software that speeds generation of CUDA-ready code 22
Look at CUDA-based work delivered as a service 22
Recommendations for vendors 22
Find ways to facilitate migration of in-house applications 22
Application vendors should find ways to exploit GPGPU 22
Alternative views 22
Appendix 23
Definitions 23
Brownian motion 23
Cache 23
Level 1 (L1) cache, Level 2 (L2) cache 23
Central processing unit (CPU) 23
Differential equation 23
Partial differential equation (PDE) 24
Stochastic differential equation (SDE) 24
Fourier transform/fast Fourier transform (FFT) 24
Finite difference method 24
Gilbert-Johnson-Keerthi (GJK) 24
GigaFLOPS 25
GPGPU/GPU computing 25
High performance computing (HPC) 25
Monte Carlo method 25
Single instruction, multiple data (SIMD) 25
Stream processing/streaming multiprocessor (SM) 25
Thermal design power (TDP) 26
Ask the analyst 27
Further reading 27
Methodology 27
Disclaimer 27
List of Figures
Figure 1: Comparing peak achievable floating point operations per second of GPUs (Nvidia and AMD/ATI) and CPUs (Intel) 8
Figure 2: CUDA Fermi: high level view 10
Figure 3: CUDA Fermi: streaming multiprocessor view 11
Figure 4: CUDA Fermi: CUDA core 12
Figure 5: AMD's vision of the evolution of HPC's processor infrastructure 19
Figure 6: Global financial markets technology spend ($bn), 2007-14 20
[Inhaltsverzeichnis ausblenden]
