What topic to choose for a master's thesis (software engineering?

Question

GPGPU

From the requirements justification of relevance, simple research, novelty.
Bachelor's degree - software for visualization of mathematical models on opengl.
But everything is bad there and in any case everything needs to be redone (with shaders and blackjack).
1) I would like to take the theme of the development of a self-written "pseudo-game" engine, but here the relevance suffers, because no one needs it for nothing. Those. if you do, then you need to name and justify it in a completely different way, in relation to some area (advise).
2) Distributed computing GPGPU. I don’t understand them at all, but presumably there is a teacher in the department who fumbles in CUDA. CUDA itself scares off platform dependence. The direction is also widespread and a narrower topic is needed (advise). Will have to think about research.
3) Because I work as a system administrator (I look after CDN). There is an option to research "methods of optimizing websites". Pretty boring I think. But as a topic, it is the most detailed, and I think it will not be difficult to attract relevance and research and novelty here.
4)Research protocols for real-time video transmission. It's also a bit boring + some protocols are closed (rtmp) + most of the sources will be in English.
5) The Vulcan API is currently gaining popularity. His research would be of interest to me. Its novelty is scary :) There are still few sources (however, there are not many of them for opengl 3-4), and bugs have not been fixed yet. Nevertheless, Hello Words for thousands of lines of code already exist.
Advise your topics or tell me how to present these to me. I still have about 1.5 years of time, so I can afford to take a not too complicated topic and figure it out from scratch.

Answer 1

The second option is clearly not the best, because. only the lazy did not write about "kudu". Basically, mage. the dissertation involves research (i.e. the program is not required, the theory is more important), so I would dismiss the first option as well. The third and fourth is it. Fifth only when compared with similar technologies.

Answer 2

I can only say about CUDA. Platform dependence here is apparent. It's just that there is CUDA, there is OpenCL, there is Intel Xeon Phi. Everything. According to the totality of parameters Where is the best now, but ... everything can change.
The essence there is only in the wild parallelization - into thousands of threads.
If you find an interesting and useful task that you haven't parallelized yet, then...