Daniel Shawul wrote:I have absolutely no idea about GPUs so I would take your word for it. But it would be a lot more interesting if someone tried and failed.
Also it seems progress have already been made on parallel alpha-beta on SIMD machines, from the link Dann provided. http://www.mpi-sb.mpg.de/~sanders/papers/gamet.ps.gz .
I've not seen _any_ SIMD alpha/beta algorithm that was worth anything. Doesn't mean one doesn't exist, but alpha/beta as we currently know it is just not suited to SIMD-type hardware.
Best I can tell from the paper cited above is it deals with synthetic trees, not a real chess search with extensions, reductions and other things that make this kind of search very problematic...
Zach Wegner wrote:You guys are all forgetting that this is the Larrabee, so you don't even have to bother doing SIMD (though it would help).
I am explicitly talking about SIMD. Only. That was my original comment in fact... But even without that, the memory issue looks daunting on a GPU-based board.
Daniel Shawul wrote:Who is ? Recently I attended a conference on fluid simulation and GPGPUs which made me very curious as to their use in chess. Some guy wrote a NS solver which outperforms the state of art commercial CPU codes 100X or so.. Some people are just amazing !
An NS Solver hundred times faster than commercial software?
Do you have some further online info on this, like conference web site etc.?
Disclaimer : I heard it during a keynote speech so don't take my word for it. 100x could be 100chips I am not sure but it was definitely shocking for me to learn of it.
Application of LES-based Model to Wind Engineering - Implementation of Meteorological effects
8:40 - 8:50
Discussion
8:40 - 9:10
Dr. Patrick Dreher presently serves as RENCI’s chief domain scientist focused on distributed computational systems and cloud computing. He is leading efforts to expand and integrate these computational systems with storage and networking to form a coherent cyberinfrastructure that supports the work at RENCI, its engagement sites throughout North Carolina and nationally.
Cloud Computing: Cloud Computing has arrived as a new technology paradigm for computation and storage. This talk will introduce this new distributed IT computation and storage option and give a high level summary as to how and where to best use this technology as part of the mix of resources to support a research program.
9:10-9:30
Dr. Rob Fowler is the Director of High Performance Computing at the RENCI. His primary technical interests in High Performance Computing are in the area of performance measurement and analysis where architectures, operating systems, and compilers interact.
Commodity Multi-Core Computers and Technical Computing: The past few years has seen a radical shift in computer design from ever faster single-threaded processors to highly concurrent multi-core and multi-thread chips. Driven by Moore's law and power considerations, this trend will continue. Meanwhile, the capacity of memory chips continues to grow rapidly, while speed and on-chip concurrency have evolved much more slowly. These trends affect the performance of technical calculations (science and engineering) on recent hardware. There are implications for future computing systems, from the laptop to supercomputers, for technical computing.
9:30-9:50
Dr. Dinesh Manocha is a Distinguished Professor of Computer Science at UNC Chapel Hill. His research group has been working on designing efficient GPU-based algorithms for geometric, scientific and database applications for more than 12 years. Some of the GPU-based utilities developed by his group are widely used in academia and industry.
Achieving High Throughput from GPUs for Scientific Applications: For years the performance and functionality of graphics processors (GPUs) has been increasing at a faster pace than Moore's Law. The latest GPUs consist of more than 3 Billion transistors and can offer a few TFlop of peak performance. They consist of tens and hundreds of cores, offer high memory bandwidth and have different programming model and the underlying architecture. GPUs as many-core accelerators have different applications, including scientific computing, database computations, sorting and geometric algorithms. New methods are being developed that could exploit the architectural features and programming limitations of the GPUs.
Daniel Shawul wrote:Disclaimer : I heard it during a keynote speech so don't take my word for it. 100x could be 100chips I am not sure but it was definitely shocking for me to learn of it.
Application of LES-based Model to Wind Engineering - Implementation of Meteorological effects
8:40 - 8:50
Discussion
8:40 - 9:10
Dr. Patrick Dreher presently serves as RENCI’s chief domain scientist focused on distributed computational systems and cloud computing. He is leading efforts to expand and integrate these computational systems with storage and networking to form a coherent cyberinfrastructure that supports the work at RENCI, its engagement sites throughout North Carolina and nationally.
Cloud Computing: Cloud Computing has arrived as a new technology paradigm for computation and storage. This talk will introduce this new distributed IT computation and storage option and give a high level summary as to how and where to best use this technology as part of the mix of resources to support a research program.
9:10-9:30
Dr. Rob Fowler is the Director of High Performance Computing at the RENCI. His primary technical interests in High Performance Computing are in the area of performance measurement and analysis where architectures, operating systems, and compilers interact.
Commodity Multi-Core Computers and Technical Computing: The past few years has seen a radical shift in computer design from ever faster single-threaded processors to highly concurrent multi-core and multi-thread chips. Driven by Moore's law and power considerations, this trend will continue. Meanwhile, the capacity of memory chips continues to grow rapidly, while speed and on-chip concurrency have evolved much more slowly. These trends affect the performance of technical calculations (science and engineering) on recent hardware. There are implications for future computing systems, from the laptop to supercomputers, for technical computing.
9:30-9:50
Dr. Dinesh Manocha is a Distinguished Professor of Computer Science at UNC Chapel Hill. His research group has been working on designing efficient GPU-based algorithms for geometric, scientific and database applications for more than 12 years. Some of the GPU-based utilities developed by his group are widely used in academia and industry.
Achieving High Throughput from GPUs for Scientific Applications: For years the performance and functionality of graphics processors (GPUs) has been increasing at a faster pace than Moore's Law. The latest GPUs consist of more than 3 Billion transistors and can offer a few TFlop of peak performance. They consist of tens and hundreds of cores, offer high memory bandwidth and have different programming model and the underlying architecture. GPUs as many-core accelerators have different applications, including scientific computing, database computations, sorting and geometric algorithms. New methods are being developed that could exploit the architectural features and programming limitations of the GPUs.
Presenter: Kurt Landrus
Authors:
Kurt Landrus, University of Arkansas Department of Computer Science & Computer Engineering, USA
R. Panneer Selvam, University of Arkansas Department of Civil Engineering, USA
Most of the time dependent phenomena in wind engineering need high performance computing. The current parallel computing techniques in cpu platform achieves efficiency of 10 to 20% and the cost of the system vary from $50K for 40 processors to close to a million dollar for 500 or more processors. Recent graphics card which have 240 processors that costs around $300 can be used for parallel computing. The overall system itself costs from $1000 for one GPU card to $5,000 with 4 cards and one can achieve speed up in the range of 100. This is an economical and viable computing one can do in their own personal super-computer. This computing is called GPU computing.
In this work, the existing computer model [1] to study flow around bridges will be considered as a bench mark study to use GPU computing. We will modify our existing 2D model into GPU system. The computer model is based on finite element method (FEM). In GPU computing the CUDA toolkit is utilized to compile the programs for the GPU. We will also look at the OpenCL cross-platform standard on multiple GPUs from different Vendors and compare the performance and power efficiency of the FEM model on the GPU architectures with CPU versions for simulation of Wind engineering applications. The challenges in GPU computing will be discussed in the full paper.
REFERENCES
[1] R.P. Selvam, S. Govindaswamy, H. Bosch, Aeroelastic analysis of bridges using FEM and moving grids, Wind & Structures 5 (2002) 257-266.
Presenter: Kurt Landrus
Authors:
Kurt Landrus, University of Arkansas Department of Computer Science & Computer Engineering, USA
R. Panneer Selvam, University of Arkansas Department of Civil Engineering, USA
Most of the time dependent phenomena in wind engineering need high performance computing. The current parallel computing techniques in cpu platform achieves efficiency of 10 to 20% and the cost of the system vary from $50K for 40 processors to close to a million dollar for 500 or more processors. Recent graphics card which have 240 processors that costs around $300 can be used for parallel computing. The overall system itself costs from $1000 for one GPU card to $5,000 with 4 cards and one can achieve speed up in the range of 100. This is an economical and viable computing one can do in their own personal super-computer. This computing is called GPU computing.
In this work, the existing computer model [1] to study flow around bridges will be considered as a bench mark study to use GPU computing. We will modify our existing 2D model into GPU system. The computer model is based on finite element method (FEM). In GPU computing the CUDA toolkit is utilized to compile the programs for the GPU. We will also look at the OpenCL cross-platform standard on multiple GPUs from different Vendors and compare the performance and power efficiency of the FEM model on the GPU architectures with CPU versions for simulation of Wind engineering applications. The challenges in GPU computing will be discussed in the full paper.
REFERENCES
[1] R.P. Selvam, S. Govindaswamy, H. Bosch, Aeroelastic analysis of bridges using FEM and moving grids, Wind & Structures 5 (2002) 257-266.
Thank you, Daniel. I have seen the number 100. I have checked abstracts with NS solvers, but should have checked GPU software.
Recently I attended a conference on fluid simulation and GPGPUs which made me very curious as to their use in chess. Some guy wrote a NS solver which outperforms the state of art commercial CPU codes 100X or so.. Some people are just amazing !