Supercomputers To Move To Specialization?

lucasw writes "The Japan Earth Simulator outperformed a computer at Los Alamos (previously the world's fastest) by a factor of three while using fewer, more specialized processors and advanced interconnect technology. This spawned multiple government reports that many suspected would ask for more funding in the U.S. for custom supercomputer architectures and less emphasis on clustering commodity hardware. One report released yesterday suggests a balanced approach."

Comments

Re:This greatly surprises me

[FullyIonized]

And I'm surprised to hear that you are surprised since fluid modeling is one of the applications that do very well with the vector processors that the Earth Simulator uses. I attended a lecture by Dr. Sato, head of the Earth Simulator, who stated that the best application usage was 65% peak usage (the theoretical peak which assumes that the processor always has data to crunch and no branches) and the average was 30% of theoretical peak. By contrast, typical fluid-like codes on current U.S. machines typically get less than 10% of peak usage if they have any type of implicitness (currently the magnetohydrodynamics code I use gives about 6% usage on an IBM SP that is #5 on the Top 500 supercomputer list).

I get tired of seeing figures that compare peak flop rates and then don't mention that actually code usage isn't keeping up at all. The Japanese (and Europeans who are allowed to buy NEC machines) are absolutely spanking the US when it comes to fluid codes (for climate modeling for example) and it is largely because they are using vector machines with their old highly optimized Fortran (or High Performance Fortran) codes. The MPP revolution in the U.S. has been manna for the CompSci community, but has set the computational physics community back by 10 years (except for those lucky bastards with embarrassingly parallel jobs).

I would give up an unnecessary body part for an Earth Simulator.