Architects have a long history of designing computers with high peak performance that is difficult for scientists and programmers to achieve. ASCI supercomputers that achieve less than 10% of peak are just the latest examples, which go back to the Illiac IV of the1960s.
Examples of recent decisions that increase peak performance AND the complexity of programming include out-of-order message delivery, pattern-sensitive interconnection networks, relaxed write consistency, very deep pipelines and correspondingly complicated branch predictors, opaque and complex prefetch engines, deeper and non-hierarchical cache levels, cluster of SMPs, limited-accuracy floating-point arithmetic, and so on.
I'll start the talk with four guidelines on how to follow the traditional path of HURTING the productivity of our scientific colleagues. This will be followed by comments on old vs. new conventional wisdom on computer technology and for scientific programming. The last part of the talk will give a half-dozen examples of how computer architecture can help scientific productivity, in case audience members are interested in this radical, unconventional goal.
I'll conclude with an overview of a new multi-university intiative to build a low-cost, research-oriented, massively parallel processor called RAMP, for Research Accelerator for Multiple Processors.Architects have a long history of designing computers with high peak performance that is difficult for scientists and programmers to achieve...all »Architects have a long history of designing computers with high peak performance that is difficult for scientists and programmers to achieve. ASCI supercomputers that achieve less than 10% of peak are just the latest examples, which go back to the Illiac IV of the1960s.
Examples of recent decisions that increase peak performance AND the complexity of programming include out-of-order message delivery, pattern-sensitive interconnection networks, relaxed write consistency, very deep pipelines and correspondingly complicated branch predictors, opaque and complex prefetch engines, deeper and non-hierarchical cache levels, cluster of SMPs, limited-accuracy floating-point arithmetic, and so on.
I'll start the talk with four guidelines on how to follow the traditional path of HURTING the productivity of our scientific colleagues. This will be followed by comments on old vs. new conventional wisdom on computer technology and for scientific programming. The last part of the talk will give a half-dozen examples of how computer architecture can help scientific productivity, in case audience members are interested in this radical, unconventional goal.
I'll conclude with an overview of a new multi-university intiative to build a low-cost, research-oriented, massively parallel processor called RAMP, for Research Accelerator for Multiple Processors.«
Download is starting. Save file to your computer. If the download does not start automatically, right-click this link and choose "Save As". How to get videos onto the iPod or PSP.
