Make it faster.
Prove it on real hardware.
A host publishes a task and a baseline. You make it faster on your own machine. hpcbench checks the answer, measures the speedup against the baseline in the same session, and ranks it.
Your code runs on your machine. hpcbench never executes it.Three moves. Nothing hidden.
Pick a problem
Read the task, the data and the score rule on one page. Fork the baseline — it already compiles and passes.
Make it faster
Change the representation, the algorithm, the parallelism — anything. An untimed prepare step builds your layout; only solution is on the clock.
Prove it
Run it on your machine. The baseline is timed beside it, the result is checked against a reference digest, and the ratio goes on the board.
A laptop, a login node and a cluster are not comparable. A ratio on the same machine is.
Every run measures the baseline in the same session on the same hardware. Your score is baseline ÷ yours, combined as a geometric mean across the sweep — so a win at one size never papers over a loss at another — and then across scoring machines. A wrong answer scores zero, however fast.
Anyone with a kernel that needs to be faster and a way to check it is right.
Engineering teams putting a hot loop out to the crowd. Research groups crowdsourcing a reference implementation. Organizers running an optimization challenge. You bring the task, the baseline and the reference; hpcbench brings the measurement discipline and the board.
Host a competition