High performance compute techniques are basically used whenever you have code that runs for 20+ hours at a time. Things like Raytracers, supercomputers, database-analysis, physics simulations, chemical simulations, weather simulations, etc. etc.
And I'm not talking about Webservers (which stay up for days, but the compute problem only happens over the course of ~1 second or so). Consider that when 3d movies are made, they hit the "Render" button and then wait 1 or 2 weeks for the movie to be made. Even if they have an entire cluster of computers calculating the scene... it takes a LOT of compute power to make movies.
If you eek out a 10% improvement, your render-time drops from 10-days to 9-days. Alternatively, you can have 10% fewer computers to get the same job done. Or in the case of scientists (ex: a Car engineer who is running a FEA computer-simulated car crash), your models can be more detailed. (Ex: Instead of modeling the engine as a perfectly elastic sphere, maybe you can model the engine as a more appropriate shape in the simulated car-crash)
There are a few exceptions. High-frequency traders also count milliseconds for their speed of execution. I see some HPC work coming out of HFT companies. But for the most part, highly-optimized code is specialized to the communities who still wait days (or even weeks) for the results from modern computers.
Comments
High performance compute techniques are basically used whenever you have code that runs for 20+ hours at a time. Things like Raytracers, supercomputers, database-analysis, physics simulations, chemical simulations, weather simulations, etc. etc.
And I'm not talking about Webservers (which stay up for days, but the compute problem only happens over the course of ~1 second or so). Consider that when 3d movies are made, they hit the "Render" button and then wait 1 or 2 weeks for the movie to be made. Even if they have an entire cluster of computers calculating the scene... it takes a LOT of compute power to make movies.
If you eek out a 10% improvement, your render-time drops from 10-days to 9-days. Alternatively, you can have 10% fewer computers to get the same job done. Or in the case of scientists (ex: a Car engineer who is running a FEA computer-simulated car crash), your models can be more detailed. (Ex: Instead of modeling the engine as a perfectly elastic sphere, maybe you can model the engine as a more appropriate shape in the simulated car-crash)
There are a few exceptions. High-frequency traders also count milliseconds for their speed of execution. I see some HPC work coming out of HFT companies. But for the most part, highly-optimized code is specialized to the communities who still wait days (or even weeks) for the results from modern computers.