When HPC will not be the HPC you remember

Just read the transcript of what sounded like an excellent talk by Greg Pfister about the next 20 years of HPC. Here are some of the key points of his talk

• Computing will become cheaper, but not necessarily much faster per processor
• There will be democratization of at least some HPC. In other words with faster processors and accelerators, we might all have access to some sort of TeraFLOPS device
• Computing will be done all over the place, with a lot being done in the cloud. I am not quite sure I get what Greg was aiming at with his section on garbage computing, but my guess is that the cycles we’ll consume might not be the highest quality cycles but they’ll get the job done
• You will be billed by how much power and bandwidth your computation consumes, not by processors and memory used
• The future of computing is not massive parallelization of the kind most HPC people care about. The future of computing is data intensive and those problems are going to be solved by embarrassingly parallel approaches

I agree with much of what’s written there. Computing is changing before our eyes. It’s not about FLOPS anymore, but efficiency; the actual cost of getting a particular job done. It’s about frameworks and techniques that optimize resources and understand failure as you scale to ginormous quantities of data. The thing is, I don’t know about 2020, but I do know about the next five years, and outside of very specialized problems, our approach to problem solving and our understanding of the worth of a compute cycle is going to change dramatically, and it’s going to be a lot of fun seeing that happen.