Yeah, I know, just like it has been the past 30 years, but for general purpose cores it’s broken now, actively, not-looking-like-it-catching-up-again broken. I’m writing this on a i5-540M (yes, the one I inserted myself to replace a A6200), a two-die monstrosity with separate GPU on it (though a crazy good deal, 2400+ passmark for $20) pressed at 32 nanometers (cue patting my laptop gently like a trusty workhorse – me and connor been through a lot) That was 2010 level tech. 22nm in 2013 was a little late, but not *that* late, and while the original law said “transistors per square inch” it’s usually more broadly used as “transistors swedes living in kansas use for daily operations”. Or (possibly) “transisotrs you can sort of expect to use in your average general purpose CPU”. Either way, bigger dies with better cooling mostly made up the difference speed wise. This year, people are pressing 14 nm, and it’s a big deal, but we *should* be pressing 5-6 nm. Intel and nvida are doing some testing at that level, but that’s “wonder if perhaps we could”, not something inching closer to production. By normal rates, we would now be kind of over 5 nm – like I’m supposed to be buying your 5 nm CPU stuff at garage sales after the HDD failed, not reading about how it’s pretty much established that it’s not a physical imposibillity.
So single thread is irreperably not catching up, which is understandable – there’s no particular reason to belive it would. But what about the fancy throughput driven kilocore stuff? I said “nVida” out loud earlier, what gives? Well, they’ve kept up, and they might just swing it. Now that Intel seems at least partially over continually shooting themselves in the foot and covering their ears going “LALALALALALA We don’t see any demand for smaller massive multicore processors” we seem marginally closer to perhaps establishing some sort of standard for computational throughput jobs and some measumrent of device cababillity for said jobs (How smoothly does this http://nothisispatrik.com/wobblebrot.htm wobbling mandelbrot operate on your machine? You know what would make it faster and smoother? You’d have to.. actually no one knows – there’s no reasonable spec to quote for massivley paralell operations besides “get a better GPU. And by “better”, I mean.. I have no idea what I mean by it”). One of Sarahs friends (unprovoked) pointed out that if you chained a pile of Raspis together and only used the GPU part, you could build a fairly cheap supercomputer, so the next generation isn’t missing what’s happening here, but most GPUs are way locked down and extremely propriatary.
So ironically, I think if we’re going to survive this itteration (because surely they world will stop spinning if we’re not doubling everyt two years ;-)) the solution is probably going to have to be political. GPU people are going to have to unglue thier cards from the vest and let us play, we’re going to have to establish a sane way to quote “dick size” (compare to camrea megapixels) so that people with money can throw tremendous amounts of it at being the best. CL is a good start, but we still need metrics, we still need higher level access, and we *might* need a killler app (although really I think that will appear organically).