[Today, Associate Editor Erico Guizzo fills us in on supercomputer pioneer Cray's new technology strategy.]
Last week, Cray, the supercomputer maker, announced a new long-term strategy. Cray believes the future of high-performance computing is integration. The company plans to combine its different supercomputing technologies into a single system that will adapt to the user's needs. Cray's buzzword for this is "adaptive supercomputing."
Cray, based in Seattle, says that today supercomputer users in government, industry, and academia are working with ever more complex models; some of these applications run well on systems that use conventional scalar processors such as Intel's Pentium or AMD's Opteron, but others require more specialized processors, called vector processors, that handle calculations in a more parallelized way. And yet others benefit from more esoteric processing tricks, such as multithreading and acceleration with FPGA (field-programmable gate array) chips.
Today, users need to write the programming code for an application—say, a climate model simulation or an aircraft aerodynamics analysis—accordingly to the supercomputer type that will be running the program. "Basically, if you buy a computer that looks like a hammer, then you have to make all of your applications look like nails. And maybe some of them want to be screws," Cray's CTO Steve Scott told me.
Cray's solution to this problem is integrating scalar, vector, multithreading, and FPGA technologies into a single machine. Users will write their programs in a standard parallel processing language just as they do now, but a special compiler designed by Cray will analyze the code and prepare it to crunch data using specific parts of the machine; some calculations will be sent off to the scalar processors, others to the vector ones, and so on.
Scott explains: "Many users actually spend more money now on programmers than they do on the machines. Software is more expensive than hardware. And so the idea behind adaptive supercomputing is take one of the hardest things for programmers off the plate: that is, don't make the programmer modify the code to try to fit it onto a particular style of architecture, which it might not be well suited for, let the programmer program in the natural way and then have the machine adapt rather than the programmer."
This integration vision will evolve in a series of steps. Initially, Cray wants to create an integrated programming environment, but the processors will exist in separate boxes in a blade architecture. The second phase involves a more highly integrated system, with the processors sharing a common memory system and interconnections. The final phase is an even more highly integrated machine with processing nodes that can decide which type of processing to run.
It's still a question for debate on whether the strategy will succeed. Among the issues Cray will need to address is cost. Cray's competitors like IBM and SGI seem to prefer relying on simple building blocks from which they can scale up large machines. By making its blocks more complex, Cray can end up with significantly more expensive machines.
Scott says that the company "is really in a unique position based on our design experience and intellectual property to pull this off." He adds that at least half of the hardware and software engineers, or about 200 people, in the company are working directly on this vision. "We're pretty excited about this," he says. "I think it's really the inevitable path forward."
<a href="http://hometown.aol.com/cup2006fifa/world-cup/websites-listings-and-reviews-for-world-cup.html">websites listings and reviews for world cup</a>
<a href="http://hometown.aol.com/cup2006fifa/world-cup/top-world-cup-2006-sites.html">top world cup 2006 sites</a>
<a href="http://hometown.aol.com/cup2006fifa/world-cup/todo-el-mundial-2006-en-terra.html">todo el mundial 2006 en terra</a>
CRAY'S NEW WAY
[Today, Associate Editor Erico Guizzo fills us in on supercomputer pioneer Cray's new technology strategy.]
Last week, Cray, the supercomputer maker, announced a new long-term strategy. Cray believes the future of high-performance computing is integration. The company plans to combine its different supercomputing technologies into a single system that will adapt to the user's needs. Cray's buzzword for this is "adaptive supercomputing."
Cray, based in Seattle, says that today supercomputer users in government, industry, and academia are working with ever more complex models; some of these applications run well on systems that use conventional scalar processors such as Intel's Pentium or AMD's Opteron, but others require more specialized processors, called vector processors, that handle calculations in a more parallelized way. And yet others benefit from more esoteric processing tricks, such as multithreading and acceleration with FPGA (field-programmable gate array) chips.
Today, users need to write the programming code for an application—say, a climate model simulation or an aircraft aerodynamics analysis—accordingly to the supercomputer type that will be running the program. "Basically, if you buy a computer that looks like a hammer, then you have to make all of your applications look like nails. And maybe some of them want to be screws," Cray's CTO Steve Scott told me.
Cray's solution to this problem is integrating scalar, vector, multithreading, and FPGA technologies into a single machine. Users will write their programs in a standard parallel processing language just as they do now, but a special compiler designed by Cray will analyze the code and prepare it to crunch data using specific parts of the machine; some calculations will be sent off to the scalar processors, others to the vector ones, and so on.
Scott explains: "Many users actually spend more money now on programmers than they do on the machines. Software is more expensive than hardware. And so the idea behind adaptive supercomputing is take one of the hardest things for programmers off the plate: that is, don't make the programmer modify the code to try to fit it onto a particular style of architecture, which it might not be well suited for, let the programmer program in the natural way and then have the machine adapt rather than the programmer."
This integration vision will evolve in a series of steps. Initially, Cray wants to create an integrated programming environment, but the processors will exist in separate boxes in a blade architecture. The second phase involves a more highly integrated system, with the processors sharing a common memory system and interconnections. The final phase is an even more highly integrated machine with processing nodes that can decide which type of processing to run.
It's still a question for debate on whether the strategy will succeed. Among the issues Cray will need to address is cost. Cray's competitors like IBM and SGI seem to prefer relying on simple building blocks from which they can scale up large machines. By making its blocks more complex, Cray can end up with significantly more expensive machines.
Scott says that the company "is really in a unique position based on our design experience and intellectual property to pull this off." He adds that at least half of the hardware and software engineers, or about 200 people, in the company are working directly on this vision. "We're pretty excited about this," he says. "I think it's really the inevitable path forward."