I’ve personally had so many problems with hardware OpenGL over the years I’ve always found myself ending up just using software (CPU + RAM) rendering.
As far as formats:
The less compression the quicker it decodes. But if your RAID can’t keep up then some lossless or extremely mild lossy compression might be in order. An extremely broad overview of this will be in an article I wrote for showreel magazine (I think July/August), however, effectively all you need to recognize is that you will hit a bottleneck somewhere in either disk performance or decode speeds, balancing that for maximum return could see some improvements in performance. But overall I’ve found that the “File In” node of any composite is usually an extremely small % of the overall render time per frame. (Somewhere around 1 second per frame. Obviously unnacceptable for real time, but compositing has never really been real-time.)
Image sequences are significantly faster to decode than movie files is another general good rule.
Be wary of involving Network Attached Storage and not ensuring it;s up to the riggors of high bandwidth random access reads and writes.
To some degree it comes down to a bit of trial and error on what you want to do. If it’s going to be a lot of 3D motion graphics with fancy DOF etc… throw more of your money at processing.
If you want dozens of layers of 4k uncompressed DPX files. Make sure you aren’t going to choke at the storage end.
P.S. don’t forget data recovery into your system cost whether that be redundant systems or an offline backup system.
So pretty much just the obvious stuff, I wouldn’t worry a whole lot… unless you’re building this yourself in which case: NEVER BUY INEXPENSIVE RAM! NEVER BUY INEXPENSIVE POWERSUPPLIES! Two rules I’ve established over the years after a half dozen unstable machines. Bad RAM and bad power supplies are the bane of system stability.