I’ve ported government Fortan code from cray to desktop in the past, and found bugs in two compilers along the way. But I won’t even look at this. Partly because my first experience was back when if you wanted more than 1M of memory on a desktop, you had to go Mac, then I did it again a couple years later on a 386. But also while there might be some fundamental issues in code optimization, I don’t think it is terribly likely. I suspect that the real gains will be in optimizing for their compiler and their hardware capabilities.
Whoever figures out how to make it run on CUDA is going to win.
Edit: or maybe not. They seem to want it to run on a specific existing supercomputer.
This isn’t about optimizing for hardware, but rather looking at the code and optimizing for algorithm. It sounds like they are look for someone to look at their code/algorithm and say you can go from X^2 to LOG(X) by changing your code to this here.
Whoever figures out how to make it run on CUDA is going to win.
I hate fortran so a generous donation to the Vintage Computer Committee on the winner’s behalf is greatly appreciated. 
FORTRAN compilers are the most advanced in the ability to vectorize computation. Anybody remember Convex in Richardson? It was acquired by HP. They specialized in relative inexpensive supercomputers.
The winner will be someone who can either make a better compiler or find a faster algorithm for what they are doing.
I read through most of the thing and they seem to want people to focus on three aspects:
Improve the multi-mesh processing speed (a cool twist on FEM that allows meshes with different shapes and coordinate systems to work in the same simulation – convenient where bodies of different shapes join)
Improve Navier-Stokes calculation efficiency (ie math)
Improve Navier-Stokes computational efficiency (ie compiler tricks)
I worked at Convex for eight years. I managed the Operating Systems, Networking and Utils group. We had an awesome compiler group and yes Fortran was the veritable forte of the compiler. It both parallelized and vectorized Fortran code to work on our fine grained parallel Super Computers. It was in fact, why HP bought Convex. There was a compiler architect there who was brilliant. His name was Steve Rowan. He was always squeezing those last few cycles out of the cpu clock time. It was fun being in the OS group when he would need some small tweak to make the compiled code run just a little faster. Man was it fun working there.
We were famous for taking “dusty decks” of Fortran code, which was originally Hollerith punch card code, and compiling it with our compilers. Organizations like NSA, Duke, Ford, GM, Bayer and DoD bought dozens of machines just because it could take this code and make it run very fast on our machines.
It was hands down the best job I ever had!
Wow, small world. I wanted to get into that compiler group, but every time I inquired, the openings were in the OS group. I probably should have interviewed anyway…
Some of the OS team would find their way to the compiler group from time to time. Getting to work at Convex as a software engineer was challenging for some, too. We had coding tests which the candidate had to do on a whiteboard in front of your peers. (You could code in any language as long as you could explain what was going on.) We started with easy bit twiddling problems then on to the dining philosopher problem (semaphore management of finite resources) which can be used to manage shared memory in a paging environment. We didn’t care if you had roller blades on when you interviewed as long as you could code. Everything was C back then and speed was everything so algorithms optimized for speed were highly sought after. We even supported research in algorithms at Rice University in hopes of further optimizing our machines.
We recruited heavily at Carnegie Melon, Univ. of Texas, Texas A&M and Purdue among others. One time a PhD student from UT interviewed for a job on our performance team. When we asked him to code he declined saying he was a PhD theory guy and he didn’t code any more. We sent him packing. Everyone at Convex in the software engineering group coded including the managers like me. It was a prereq for working there.
Unfortunately, the OS left a bit to be desired.
(A recovering SPP-1600 and HP V sysadmin from decades past.)
I had nothing to do with the SPP or HP series machines. The ones I worked on were the original C-series.
The SPP work was just beginning when I left.
I spent 27 years at Control Data Corporation. The most fun part was the 6600 and 7600 era when they dominated supercomputers. About a year was spent at Mostek and much of the time was converting and optimizing Spice Fortran programs. Every design engineer seemed to come on board with his own personal highly modified copy of Spice.