Since these are mobile processors your understanding is incorrect. What you've described is what AMD did on the server & desktop side. But on mobile, to hit the power efficiency targets, AMD stuck with the single monolithic die. And it's a decently large one, at that: https://www.anandtech.com/show/15381/amd-ryzen-mobile-4000-m...
So no AMD wouldn't yield better on Intel's 10NM+, not in this case. In the desktop usage they would, as their desktop CPU dies are way smaller than their mobile CPU dies. But the mobile die sizes are fairly equivalent between AMD & Intel.
> The fact that Intel's only new releases on 10nm are <=6 core low power products suggests their yields are still poor.
Their argument was that, had they followed AMD's approach, they would be able to produce desktop and server chips despite being yield-limited to "<= 6 core" per chiplet.
Wow. Is this a relatively new phenomenon, or something that's been the norm in processor mfg for a long time? I know that the Ks were usually just binned/higher tested chips so they'd unlock those and charge a premium for them, but the idea that they're disabling parts of the die and separating the chip offering that way is crazy to me. Are the parts being disabled redundant or are they reducing the instruction set space? I don't know anything about how any of that would work.
Something that's different between software and hardware such as chips are that the chips that have things disabled are done so because those disabled parts actually don't work. If you only yield 2 out of 4 cores, it's easier to just make it a dual-core CPU rather than throw the chip away.
So no, nothing new under the sun.
https://www.zdnet.com/article/upgrade-your-radeon-hd-6950-to...
wait, it's all just binning?
it always has been.
;)
https://en.m.wikipedia.org/wiki/Product_binning
Iirc, early Celeron were often just Pentiums with defective (and disabled) cache
That logic seems backwards. If you have poor yields you tend to favor higher-margin products (i.e. datacenter CPUs). This is a laptop CPU intended to sell at significantly lower $/mm2.
As far as chiplets: multichip packages are an ancient idea, and the industry goes back and forth on them. Both Intel and AMD shippsed multi-chip solutions way back in the day, the current age of integration is actually the anomaly. You win on yield but lose on package costs, and the decision as to which to use depends on the specifics of the market you're trying to target.
Certainly AMD would prefer to ship single-chip solutions and pocket the savings, but they can't. Likewise Intel accepts some loss of scaling because of the need to share die designs across the product line.
I say this having no idea what Intel's defect rate is right now, and I acknowledge fusing of bad sections can mitigate this a bit.
AMD did a really smart thing with their processor designs. Instead of doing a large monolithic multiprocessor like Intel (which may be able to eek out some extra performance due to optimized wire placements), they use smaller chiplets. This results in requiring no manufacturing errors on a much smaller area of the die (which becomes exponentially less probable as the area increases). Even if AMD was using Intel's 10nm process, their yields would be better because of their modular designs that get soldered together after the lithography. Intel can still bin out some errors (the 10300-10900k are all the same chip with different parts turned off due to manufacturing errors), but the less modular design likely makes this less efficient.
That's my understanding, anyway -- correct me if I'm wrong.