Thanks for the kind words, @LaFleur.
I have never ventured into annual PIT renderings (8760 images). The longest time period has been one full month (744 images), but that was with matrix-based methods when I did some testing in relation to integration of the imagebased 3-phase method in HB[+]. Another option is the imagebased DC method, which is a more stable recipe in HB[+].
The obvious benefit of matrix-based methods is that you only have to do one simulation, and the rest is matrix multiplication to generate images for each hour. While that is great, I did experience that the matrix multiplication (dctimestep) is also a slow process. Multi-core is not possible for dctimestep by default, though running dctimestep in parallel is possible, but not part of Honeybee. During testing I also experienced that dctimestep is significantly slower on Windows compared to Linux - I used Windows Subsystem for Linux, so it was still on a Windows machine. The comparison is seen in the image below - note that the 6 cores for Linux is only for rcontrib and not dctimestep.
So to answer your query, I am not sure I have a better solution, but if I was forced to do 8760 renderings I would likely opt for matrix-based methods (HB[+]) rather than rpict (HB-Legacy). And as you mention it will take up a lot of memory. I currently only have a 256 GB hard drive - I wouldn’t even make it past January!!! 
