Graphics Processing Unit (GPU) Panel: Making Your Desktop a Supercomputer?

Chair:

Prof. Dr. Kun Zhou, Cheung Kong Distinguished Professor, Computer Science Department, Zhejiang University, China

With multiple cores driven by very high memory bandwidth, today’s graphics processing unit (GPU) has involved into an absolute computing workhorse. The latest GPU contains up to 240 processing units supporting thousands of simultaneous threads, a level of concurrency that cannot be found in any other consumer platform. Although designed originally for 3D graphics applications, GPUs have become general-purpose architectures with fully-featured instruction sets and rich memory hierarchies. All these amazing power and functionality are packed into a palm-sized video card, which can be easily attached to a PC or workstation through a standard PCI-E connection. This puts supercomputing power into the hands of every scientist and engineer – your desktop can be easily transformed into a personal supercomputer.

In this session, five GPU experts from both industry and academia will share their perspectives on several aspects of GPU technologies, including hardware architecture, software tools, and applications. After their talks, we will also have some time to receive comments and answer questions from the audience.

From the hardware side, Dr. Justin Hensley from AMD will introduce the basics of graphics hardware. In particular, he will talk about how the hardware evolved into what it is now and how the hardware works in an abstract level. While NVIDIA and AMD control nearly 100% of the high-end GPU market, Intel recently presented new insight into the next generation GPU architecture: Larrabee. In a subsequent talk, Dr. Pradeep Dubey, Senior Principal Engineer from Intel, will tell us the main ideas behind Larrabee.

Programming GPUs used to require significant graphics-specific knowledge from the programmers. Recent GPU languages like CUDA and OpenCL comprise an extension to the C programming language for a minimum learning curve. From the software side, Dr. Massimiliano Fatica from NVIDIA will introduce the CUDA programming model and motivate the use of CUDA with many snippet examples from different high performance computing domains.

All these wonderful horsepower and programmability of GPUs would likely be under-utilized if your applications remain sequential. Although hardware architectures and software tools could help parallelization, human intelligence is still likely needed in designing high-level algorithms for parallel applications. From the application side, Dr. Li-Yi Wei from Microsoft will introduce some examples from the field of graphics and statistics, telling us how to parallelize some algorithms which are thought to be intrinsically sequential.

This session is targeted at IT managers who want to better understand the state-of-the-art GPU hardware and software, as well as scientists and engineers who are interested in utilizing GPUs to speed up their applications.