
How The Mighty Have Fallen...
They can't imagine competing with AMD so they just petulantly throw in the towel.
Intel's grand plan to stitch CPU, GPU, and memory dies together on a single package called an XPU is dead in the water. In a press conference ahead of this week's International Supercomputing Conference (ISC), Jeff McVeigh VP of Intel's Supercomputing Group revealed that the company's Falcon Shores platform wouldn't just be …
Indeed yes.
AMD look like they're also going to show what's possible too, when you want a no holds barred, really, really fast chip! I'd love to see the details of how AMD are meshing their CPUs, GPU and memory together (probably, a lot of cross-interconnect). I'm assuming Apple's sillicon (totally different thing of course) has a lot less mesh in the on-chip fabric. Fabric design is non-trivial - I wonder how good Apple are at it, and can they expand it? Looks like AMD have become pretty good at it, to wire all those chiplets up together.
Fujitsu are pretty good at fabric design too - they've put their Tofu interconnect into the ARM based chips the Japanese super computers run. Though in that case, it's mind bogglingly complicate (6 dimensional hypertoroid blended with a 4 dimensional hypercube, if you want an inkling of the interconnect routes in those machines).
Intel & Nvidia "GEE, THIS CHIPLET THING IS HARDER THAN IT APPEARS!"
Way back in Q1 2017, starting with a simple 4 core ccx/lego block, AMD could assemble 4/8/16 CCX into socket modules for desktop/ws/server respectively.
Even further back in 22014 or sooner, they were doing similar with 2x 2core ccx + a very impressive GPU in APUs like the 7850k.
It is now nearly q3 2023, & we have lame, same ol 8 core monolith intel, plus ecores since feb 2022 & a now delayed & now canceled APU
Nvidia is similarly unimpressive - they cobbled together their old risk cpu & gpu & gave them separate caches & a nvlink inteconnect.
In over 6 years of Infinity Fabric/chiplet omnipresence in the market, this is the best NV & Intel can manage?
amd is slaughtering intel in DC, & MI300 slaughters grace hopper.
~NV can manage a cpu/gpu combo, & intel, a perf cpu w/ 1 or 2 (sometimes useful w/ MS help) e core chiplets.
Even amd chiplets have chiplets. MI300 has ~15 (high yield & cool running) chiplets &15billion transistors.