So what else are you guys lokoing forward from Clanser??
Tuesday, December 6, 2011
Tuesday, May 18, 2010
HPC Processors Comparison
To buy an HPC system the first and the foremost task is to select the processor. I am attaching some slides which are the average results the processors are providing across various vendor platforms. The data required for these have been taken from Spec.org. Moreover the data shown here is from the systems tested after May2009.
The first three set of images show the comparison between CFP2006 results, FEA averaged results and CFD averaged results. For a fair comparison only processors with same number of cores/chip have been compared.
Comparison based on Overall Floating Point results
The first three set of images show the comparison between CFP2006 results, FEA averaged results and CFD averaged results. For a fair comparison only processors with same number of cores/chip have been compared.
Comparison based on Overall Floating Point results
8/12 cores per chip
Comparison based on FEA codes
4 Cores per Chip
6 Cores per Chip
8/12 Cores per Chip
Comparison based on CFD codes
4 Cores per Chip
8/12 Cores per Chip
Power 7 based CPUs definitely have are the best in the market at the moment, but AMD are still the cheapest providing more cores in the same price. Overall I would prefer Intel Xeon in the current processor market.
I have been an overall fan of AMD all my life, but in the current market. AMD seems to be loosing. Lets hope some day we will have an HPC under our desk to perform simulation.
Monday, May 17, 2010
HPC Reference Sites
These are some of the sites for your reference.
Third Party Sites
http://www.spec.org/
http://www.hpcwire.com/
http://www.top500.org/
http://hpc.sourceforge.net/
Vendor Sites
http://www.hp.com/techservers/index.html
http://www-03.ibm.com/systems/deepcomputing/
http://www.cray.com/Home.aspx
http://www.microsoft.com/hpc/en/us/default.aspx
http://www.linuxhpc.org/
Third Party Sites
http://www.spec.org/
http://www.hpcwire.com/
http://www.top500.org/
http://hpc.sourceforge.net/
Vendor Sites
http://www.hp.com/techservers/index.html
http://www-03.ibm.com/systems/deepcomputing/
http://www.cray.com/Home.aspx
http://www.microsoft.com/hpc/en/us/default.aspx
http://www.linuxhpc.org/
Buying a new HPC system
I am sure many of you would have brainstormed a lot on which kind of HPC system you should be buying for your FEA/CFD applications. There are a range of servers available in market, but which one is best suitable for HPC?
To start with HPC stands for High Performance Computing, for those who like a more fancy name, one can refer HPC's to an affordable Supercomputer.
What should you be looking for, when you are thinking of buying an HPC system. FLOPS or the Floating Point operations per second. We in the field of HPC are not interested in Integer calculations. So if someone is trying to sell you a system which is good in Integer calculations, you are probably being misinformed.
One should remember that common servers available in market are either to operate Java, C , C++ or some dataserver. While in HPC we require systems which are good in Fortran. One of the best source of information about these is http://www.spec.org/cpu2006/CFP2006/ . These people have benchmarked systems by solving CFD, FEA, quantum chemistry, Molecular dynamics, Linear programming, computational electromagnetics etc. So this is basically a list of all the codes which require floating point operations.
Now before we purchase any system, if we are clear that our main use is either FEA or CFD then we can be more streamlined in our outlook. Else an overall CFP2006 score should be good enough.
The scalability of CFD is usually linear, computation time is directly proportional to the number of available cores. While for FEA it is difficult to attain high scalability. Moreover FEA works better if you have more number of cores on a single motherboard, but if you are going for cluster an Infiniband is must.
I would suggest as a thumb rule that 4GB of RAM per core is much more than needed for a CFD or an Explicit simulation. But if you are doing lot of matrix inversion or eigenvalue problems, then it is better to have as much memory as possible. I have personally seen that more than 12 CPUs might sometime slow down such problems.
High processing speed is definitely good, but you will should have high cache memory. 1MB per core should be the minimum you should be looking for. Anything less than this will be only a marketing gimmick for your organization.
So before you buy, do your homework.
To start with HPC stands for High Performance Computing, for those who like a more fancy name, one can refer HPC's to an affordable Supercomputer.
What should you be looking for, when you are thinking of buying an HPC system. FLOPS or the Floating Point operations per second. We in the field of HPC are not interested in Integer calculations. So if someone is trying to sell you a system which is good in Integer calculations, you are probably being misinformed.
One should remember that common servers available in market are either to operate Java, C , C++ or some dataserver. While in HPC we require systems which are good in Fortran. One of the best source of information about these is http://www.spec.org/cpu2006/CFP2006/ . These people have benchmarked systems by solving CFD, FEA, quantum chemistry, Molecular dynamics, Linear programming, computational electromagnetics etc. So this is basically a list of all the codes which require floating point operations.
Now before we purchase any system, if we are clear that our main use is either FEA or CFD then we can be more streamlined in our outlook. Else an overall CFP2006 score should be good enough.
The scalability of CFD is usually linear, computation time is directly proportional to the number of available cores. While for FEA it is difficult to attain high scalability. Moreover FEA works better if you have more number of cores on a single motherboard, but if you are going for cluster an Infiniband is must.
I would suggest as a thumb rule that 4GB of RAM per core is much more than needed for a CFD or an Explicit simulation. But if you are doing lot of matrix inversion or eigenvalue problems, then it is better to have as much memory as possible. I have personally seen that more than 12 CPUs might sometime slow down such problems.
High processing speed is definitely good, but you will should have high cache memory. 1MB per core should be the minimum you should be looking for. Anything less than this will be only a marketing gimmick for your organization.
So before you buy, do your homework.
Sunday, May 16, 2010
CLansers
CLancers.. Freelancers in the field of CAD/CAM/CAE.
We hope that we will be able to provide you with technical help in this site. You all are welcome to comment on this site, and we will try to come up with an article on your comment.
This blog is maintained by the creators of the Cnnovate Solutions.
We intend to talk about most of the ongoing stuff in the field of CLansing.
Happy Clansing..
We hope that we will be able to provide you with technical help in this site. You all are welcome to comment on this site, and we will try to come up with an article on your comment.
This blog is maintained by the creators of the Cnnovate Solutions.
We intend to talk about most of the ongoing stuff in the field of CLansing.
Happy Clansing..
Subscribe to:
Posts (Atom)
