I've been thinking a bunch about this over the last several days, and I think there are perhaps many "parallels" with the computer industry to be explored.
A few years back, personal computers hit a performance wall. After years of marketing ever higher mhz then Ghz ratings as metrics for CPU performance, it just stopped being a relevant comparison. A host of other performance bottlenecks began to be the limiters (memory latency, disk I/O, bus bandwidth, etc). As one of those types who has actually bought CPUs from time to time, it is a pretty amazing thing to me that the range of advertised clock speeds has barely changed since 2005 - literally eons in computer time. But really this is just a manifestation of just how rapidly the focus shifted to other areas of progress - particularly the development of chips that feature multiple processing cores operating in parallel.
In the mid-1990's being able to do computations in parallel meant: 1) having access to a fancy parallel computer like the Cray t3-E (think super computer center at large university), 2) being capable of programming in either C or Fortran, and 3) having a problem that was actually amenable to the complete rethinking that parallel computations required. It was hard, and not many people did it, but the results were amazing - real time weather simulation, modeling atomic explosions, proteomic modeling etc. Fast forward to today and heck even my laptop has multiple cores. And it helps when I do sharpening in Lightroom.
My point is, that if you look around at the technology in photography today, it is clear that a similar process is well under way. Stacking and stitching shots is a pretty useful way to knock your way through many of the barriers that single-chip photography presents. For example you can pack your pixels almost as deep as you want (depending on your patience and focal range of your lens kit). By the way, isn't this one way around the diffraction limit: you just bump up the magnification factor and increase the virtual size of the sensor? You can also make focus as shallow or deep as you want it. You can boost dynamic range by multiple stops. And, of course, you can pick your field of view on the fly.
Practicing this dark art, though, reminds me a lot of the early/middle years of development in parallel processing. It has become really easy for the technically inclined, and it can produce mind-numbingly good results when it works. And all of these things can be had with cheap stuff too: a workable pano head costs under $100, and the software starts at free (but the good stuff will set you back a few hundred too). My Canon trusty old 350D seems to work fine in this application.
But then there are the limitations. Forget about moving subjects (unless you are really creative and plan). Getting the captures right can be hard too - do any cameras out there have focus bracketing? Even just just remembering to get all the combinations can be a challenge: DOF x DR x X x Y gets to be a lot to remember. And it can take real time and patience - though what's new, right? Also, though you no longer have to master the secret handshakes of Panotools, you still have to know what you are doing. Oh, and there's one other investment - processing time (or a really fast computer) to make the stacks and stitches happen.
Of course some of the limitations are not really limitations, but the challenges associated with trying to figure out what to do with all the DOF/DR/FOV that you get to work with. Getting something that doesn't look bizarre out of HDR is a real trick, and just moving a modest multi-hundred megabyte hdr panoramic file around on a computer can be a chore. The creative opportunities are significant, but there's a price.
It's all really promising, but it's also clear there's still a ton of room for development. For example, there's more natural convergence on the software end of things waiting to happen (though this seems to be progressing quickly). And on the hardware end of things too - what could be done? What about a programmable head/camera combination that could walk through the combinations automatically? What if that could do a quick snap and stitch to preview your composition? Is there a workaround for the need to swing the lens around the nodal point such that multiple lenses/bodies could be triggered simultaneously? Could multiple cheap chips/lenses be used simultaneously to break the time barrier that currently exists - i.e. really go parallel? Etc. etc. And, probably not all photographic problems can be solved with these tools - even with more development.
The economics of multiple sensor photography bear consideration in this equation. It's just a ton cheaper to go parallel with your existing camera than to invest in a medium format back. That will probably not change. My understanding is that what has made modern computer processors so cheap is that increasing transistor density has mainly delivered higher per-wafer yields. Building big sensors will likely always be more expensive because covering a larger wafer area with one chip increases the chance of pulling in enough bad pixels to result in a failed unit. If you cover the same area with eight chips and only one of them lands in the area with the failure, then you still have three chips and much less waste.
I suppose that if you could figure out how to build a camera using an array of small chips, then it could be quite cheap.
Interesting stuff - anyone else out there have thoughts like this?
Chris