the cyclical nature of moving things off the CPU die into accelerators, and then moving them back onto the CPU die as the industry rediscovers that the principle of locality exists...
Comments
Log in with your Bluesky account to leave a comment
Not even just CPU design... Everything gets decentralised then re-centralised, all up and down the software stack, in loops, forever. And it makes it all slower! But nobody seems to notice this pattern..
If I squinted my eyes and tilted my head, it looked like a Unix and software stack to me. More interesting as the things scaled through racks, spines, data centers and hubs. 🤷🏽♀️
New computation becomes commonly done in bulk -> New silicon architecture made to speed up common computation -> Modular expansion, using that architecture, available to those who need it -> Expansion adopted by enough users that devs assume people have it -> Expansion integrated into main CPU
it's the same cycle of moving stuff from the local system to The Mainframe/The Server/The Cloud and back too, fractally up and down the whole ecosystem
My “favorite” part of everything before the Pentium era is that there were no alignment pins on any of the PGA packages, so you could put the chips in any of four ways, three of which would produce smoke, and then of course some motherboards didn’t even point them the same direction
xeon phi is so cool to me because it's like an intel engineer saw the "would you rather fight one horse-sized duck or 100 duck-sized horses?" reddit thread and was like "wait, i have an idea!"
The 80387 is the reason I switched from assembly over to C. The 8087 was great to deal with and the 80287 wasn't too bad, but I just couldn't face handling the 387 in assembly.
unironically what you do with pretty much any accelerator that’s large enough (just write a soft core for fpga or embed a small risc-v processor for asics)
Comments
latency still matters.
(from 1968) https://dl.acm.org/doi/10.1145/280811.281014
I don’t understand any of the real mechanics. I suppose if things are ‘spaced out’ you’re limited by the tracks connecting them.
(I mostly know biological science)