AFAIK, that is not how it works. The cpu is not thinned after the wafer is diced. The whole wafer is polished down so there is no binning before thinning. In fact, the stacking process is likely at wafer level also. They thin the cpu wafer and then make a “reconstituted” or “carrier wafer” that contains cache chips (diced) and filler silicon. The entire wafer is stacked, bonded, and then diced. I assume that this also why it is limited to wafers made at the same facility.
When using thinned dies in the past, they were always diced first as getting a flat thinning across the entire wafer is difficult, that's what the foundry told us anyway but it probably comes down to the tools used (there are different methods of thinning) and how thin you want the substrate. I imagine the thinner the substrate you need, the more likely you are that it will have to be diced first as you also lose structural stability after thinning. You are probably right though for HVM like AMD is going for and with only a single stack, the whole wafer is thinned first.
This should work fine for binning since Milan-X should be able to absorb parts with defective cores. They will probably go down to 2 or maybe even 1 core active for special large cache per core versions like the current 72F3 which is 8 cores and 256 MB L3. That one is the full 8 CCDs with just one core active per chip. These can be a little more leaky since the core count is so low, but the rated TDP is also not that high, so there are some limits. They don’t have use for 8 core leaky die in Milan-x, but the lower core count die are mostly still usable for Milan or Milan-x parts of some kind. The 8 core die are mostly used in 32 or 64 core devices, where the power consumption would be limiting so they should have a lot of 8 core die that are not usable in Milan-x.
I don't know about a lot of leaky dies(relative to total number). Modern processes are getting pretty good at constraining process variation though I don't know the specifics of 7 nm. With the node being this mature, though, I imagine they have it pretty well under control. Obviously I'm not privy to the actual data though.
Only selling the 8 core makes sense in a lot of ways. The number of people who have a 12 or 16 core and who would be willing to upgrade to an X3D version is very small. They would probably rather have those people spend the big money later in the year for a completely new Zen 4 system. Going to dual CCD devices cuts volume in half for something that is likely already low volume. It also might cannibalize other sales if they released such a device. I would definitely consider such a 16 core for a compile machine instead of Epyc or Threadripper / Threadripper pro if it was available. If the device had come out a few months ago, then a more full product range might have made sense. A lot of things have been delayed due to covid shutdowns and such. Even regular Milan availability seemed to be a problem for a while. We are not going to get new threadrippers until the Milan backlog can be resolved.
Yes, it makes sense, I just disagreed that the reason we went from AMD implying at least a 5900x used to only a 5800x is because AMD was surprised by the Milan-X demand or by how many leaky dies they were getting through fabrication. Neither of those explanations makes sense to me. I'm still not sure on the market position though. If AMD is accurate in the +15% performance estimate then it's not much added value for someone who would shop at the 5800x levels, especially since they are likely to be GPU limited anyway. I think people who just want the best period are more likely to spend the extra money and be shopping in the 5900x-5950x levels to begin with. We'll have to wait and see what the final cost is though.
For most people, it will be wait for Zen 4 unless you need a new system right now or have an older chip that can be upgraded. It is likely not going to be worth it if you already have a 12 or 16 core. I would still consider a 5800X3D because I have dealt with board revision 1.0 before and I have an ancient desktop system at the moment with no upgrade path. Zen 4 will be new everything and expensive. The 5800X3D will be probably be the last and highest performance AM4, except some heavily threaded applications where 12 or 16 cores may still win. Waiting for Zen 4 and DDR5 might be a longer wait than expected due to the inevitable shortages.
Yes, it still offers a very good upgrade path for a lot of people. It will just come down to how much AMD will charge for it. If it creeps to close to a 5900x, I can see even people who only game going for the 5900x instead to have more cores to be more comfortable about future performance or because they do (or at least think they do) enough in parallel to gaming to justify the additional cores.
forums.anandtech.com
Facebook
Twitter