Note that they were talking about _assembly_ code paths and parent explicitly mentioned using anything else then assembly would give you more portability.
Even though the FPGA fabric might encode the solution more effectively, there are other important differentiators: clock speed and memory bandwidth. GPUs have higher clock speeds and typically better memory bandwidth (related of course).
With the higher clock speed, GPUs can well outperform FPGAs for many problems.
I guess there is some AI algorithm that does zoom as postprocessing. That AI knows the moon so it can fill in the blanks and compensate for a (relatively) crappy sensor.
So in the future, there are either cameras that can see what others have seen before, and those that can truly capture new, true, detail (true as in, without filling it with estimations)
Not only is Martin Blais' explanation very helpful, his software, Beancount, is also very suitable for any personal finance projects, especially if you have a background as Developer.
Personally, I combine beancount with fava and find it much better than e.g. GnuCash.