The last few years have been pretty far out of the norm as to what you would normally expect from endpoint computing. AMDs strong performance is due to all kinds of things from Intel fumbling hard, to work trends shifting, as well as just having a really solid product. They did really well in client side of things but you have to wonder how it would have looked if Intel wasn't so far behind and the market demand as strong as it was. You can certainly make a lot of money in the client space but its done at scale and thats not something AMD can do like Intel can.
No matter how you look at it margins are better in datacenter and HPC parts. Not only are the margins better, once you win one datacenter contract thats good for at least five years minimum as platforms are expanded and upgraded, and replicated across sites so that one initial win is pretty much guaranteed to provide future sales for years down the road.
You're not entirely wrong, but remember: even back in 2019, at the peak of Intel's 14nm supply crunch and before AMD was making much headway into the datacenter space,
their consumer branch was still outselling their datacenter branch by nearly 39% (US$ 10bn v. 7.2bn). People tend to dramatically underestimate the scale at which consumer computing operates, and overvalue datacenter sales. Yes, datacenters are a huge and high margin market, but it's also a high cost of entry and high maintenance market, which eats away at those margins. It's still very profitable, but so is the client computing market, at least for CPU makers.
IDK, if you had limited supply who would you rather say "no" to. DellEMC, and HPE or a bunch of fickle gamers?
It's not an either-or situation. The question is where one finds the balance between the two. And that's what's being criticized here.
Its fair to criticize them for the priority decisions they made but I would say the made the right business priorities. Ideally I'm sure they'd go after all of the markets including entry-level to mid-range markets that would be served by a 4 core Zen3 but I don't think they have the resources to pursue all that at the same time. Zen3 chiplets are already pretty small (compared to Alder Lake anyway) and yields are really good so while you'd get more by making a smaller chiplet and slightly better yields there is less money in each chip so there isn't much incentive to do it when your 8 core chiplet is yeidling great and selling out. Also, nobody really buys server CPUs in less than 10 core per socket configurations so a 4 core chiplet would be pretty much useless for Epyc.
And it's entirely reasonable to criticize AMD for how they've gon about this either way: I'm not beholden to their profit-only corporate logic, nor do I see any reason to accept any deeply ideological arguments towards their "main purpose being to make money" or anything of that sort. The only reason companies exist in the first place is in order to produce somehow useful products - thought this can be manipulated greatly, profits are always secondary to this in the end. This is also the basis of my criticism: I disagree completely with their choice to entirely ignore low-end CPU and APU markets for the past years.
You're also missing the point when talking about yields - the point of a smaller die wouldn't be better yields (those are already great), but simply increasing the die per wafer count. Heck, if anything, the good yields from TSMC 7nm is a
disincentive towards producing low-end chips from an 8c CCX, as there simply aren't enough defective chips for this. That's why I brought up a smaller die to begin with. A thought experiment (which is obviously very simplistic, but still useful): the Zen3 CCD is reportedly 11.27x7.43mm. Putting that into a
die per wafer calculator alongside TSMC's (old, likely superseded) 0.09/cm² defect density, gives us 692 complete dice per wafer, with ~50 defects. So, ~642
fully functioning dice, with 50 that might be salvaged with a lower core or CU count or cache size (though they may also be unsalvageable). AMD has essentially no reason to sell lower core count chips unless they just can't meet power or clock targets. So, let's assume a 4c CCX would just have the bottom four cores lopped off (it would likely entail more of a reconfiguration of the layout, but the area would still be similar). That would make it ~11.27x4.2. Plugging that back into the calculator gives us 1228 total dice with 51 defects or 1177 fully enabled dice per wafer. That's not quite a doubling, but it's still a
massive increase in output per wafer. And these dice would of course be equally usable across their entire product range, from consumer desktop to OEM desktop to low core count/high RAM bandwidth/high PCIe lane count servers. And, of course, you could still make a 32c EPYC with these if you wanted to, though it would have slightly higher latencies than 4x8 core equivalents. Designing the new die would be a significant cost, but it would give them a product portfolio much better suited to addressing a diverse market, while drastically increasing the actual numbers of chips produced. And allocating even 10% of the wafers used for producing 8c CCDs would have given them a pretty significant amount of 4c CCDs.
Of course, there are more factors than just CCDs in this, and there have been reported shortages of CPU packaging and substrate materials, hold-ups with downstream processing plants, and it's not given that they have sufficient IOD supplies for this either. They also likely banked on their single Zen3 die to hold them over till Zen4 alongside the new APUs, but this has clearly failed in light of the shortages.
But most of that is unknown and unknowable, and doesn't change the fact that I simply think AMD made the wrong judgements here, and reacted poorly to the shortages. They tried to ride things out and keep prices high while investing minimally, which is backfiring significantly against a resurgent Intel that competes across a much broader range of products. They've left the low end essentially barren for two product generations and multiple years, which is undermining the public support that allowed them to gain their current position in the first place. This might be a passable plan if you're in an entrenched position of power, but AMD isn't. They're still highly vulnerable to the much larger and wealthier Intel. And especially in light of that, but also in light of their responsibility to produce actually useful and somewhat attainable products for their customers, I think they've missed the mark.
If you need 2 x pcie 4.0 m.2 then B660 is they way to go.