Supercomputer performance leaps off the charts
Where we are with HPC, where we’re going
Some interesting data in the slides, including performance metrics of the #1, #100 and #500 systems on the Top500 list over the past ten years. When you chart out the results, you see a pattern where we just sort of plodded along until somewhere around 2006 – and then performance ramps up in hockey-stick fashion.
This pattern holds true for the top, middle-high and lowest systems tracked by the Top500. The performance growth rates are astounding. From 2001-2010, performance on LINPACK for the #1, #100 and #500 system on the Top500 list grew at compound rates of 92%, 85% and 90% respectively.
This is a much higher rate than what we’d get from Moore’s Law like advances in processors, interconnects and storage. Rich and I see hybrid systems (CPUs + GPUs or other accelerators) as the nitrous oxide injected into the engine block of HPC system design.
We talk about the costs of HPC and they’ve remained stable for the most part and drop when hybrid systems are factored into the mix. There’s some conversation about the importance of ecosystems and how they foster new technologies plus even some chatter about HPC in the clouds. Give it a listen and let us know what you think ...
Bootnote (We talk about some CAGR figures in the audio of the webcast which don’t match up to the slide set you’ll see. The CAGR calculations on the original slide set were incorrect and understated the performance growth rates quite a bit.
The numbers on the slides now are pretty close – the exact rate of growth varies depending on how you calculate it. I simply took the initial value in 2001 and found the annual growth rate that it most closely match the 2010 value, which was close enough in my mind.)