It is guaranteed NOT to stay the top tier. Also, as of the xx9x\x9xx family we have sub-tieres in the top tier. 4090 is the low/mid top tier because 4090 ti is almost certainly exist.
That is literally exactly what I said. That's the distinction between being
the top-tier SKU and
a top-tier SKU.
If you consider "what is the best" and diciding product A vs. B according to that plus you are willing to pay more only to be entitled to that "the best" treet by itself than, well, you condemn yourself into a limbo of dissapointment.
No. All purchases are a judgement of value for money, and when buying a product you cannot escape such judgements even if you explicitly don't care about value - it's there in literally every aspect of the thing. If you're paying extra for a top-tier product - which you inherently are by buying a flagship GPU - then you're putting your money and trust into a relationship with another party based on their messaging, i.e. marketing. If that company then abuses that trust by subsequently changing the basis for the agreement, then that's on them, not on the buyer.
"...then it's reasonable to expect it to stay as the best for a while"
This is a very naive, childish approche imo.
What? No.
A while is not a fixed amount of time. It is an inherently variable and flexible amount of time. That's the entire point. There's nothing naive about this, it is explicitly
not naive, but has the expectation of its end built into it. The question is how and why that end comes about - whether it's reasonable, i.e. by a new generation arriving or significant refresh occurring, or whether it's unreasonable, i.e. through the manufacturer creating minuscule, arbitrary binnings in order to launch ever-higher priced premium SKUs.
No one has or will guarantee you a time frame of "the best". Expecting such thing is way out of scope of product spec hace the gape leading to disappointment.
No it isn't. The
explicit promise of a flagship GPU is that it's the best GPU, either from that chipmaker or outright. Not that it will stay that way forever, not that it will stay that way for any given, fixed amount of time, but that it is so at the time and will stay that way until either the next generation or a significant mid-gen refresh.
But to which is own I guess, just please don`t use that disappointment to bash any company. That being bias.
Yes, of course, all criticism of shady marketing practices is bias, of course. Unbiased criticism is impossible! Such logic, much wow! It's not as if these companies have massive marketing budgets and spend hundreds of millions of dollars yearly to convince people to give them their hard-earned money, no, of course not
Seriously, if there's anyone arguing a woefully naïve stance here it's you with how you're implicitly pretending that humans are entirely rational actors and that how we are affected by our surroundings is optional. 'Cause if you're not arguing that, then the core of your argument here falls apart.
Not everything is some shitty move, the reasoning is pretty clear. The flagship GPU is so large that yields play into this, to make the volume they want to be able to sell given their market share, they can move vastly more with relatively minor defects and start to stockpile ones with no faults/good binning etc, then when enough exist to make a product viable, they'll sell it. How shitty of them.
No, not everything is a shitty move, that's true. But you're giving
a lot of benefit of the doubt here. An unreasonable amount IMO. Even if initial yields of fully enabled dice are
bad - say 70%, which is borderline unacceptable for the chip industry - and that a further 50% of undamaged dice don't meet the top spec bin, what is stopping them from still making a retail SKU from the remaining 35% of chips?
The problem is, you're drawing up an unreasonable scenario. Nobody is saying Nvidia has to choose between
either launching a fully enabled chip, or a cut down one. They could easily do both - supplies of either would just be slightly more limited. Instead they're choosing to only sell the cut-down part - which initially
must include a lot of chips that
could have been the top-end SKU, unless their yields are absolute garbage. Look at reports of Intel's fab woes. What yield rates are considered not economically viable? Even 70% is presented as
bad. And 70% yields doesn't mean 70%
usable chips, it means 70%
fault-free chips.
A napkin math example: AD102 is a 608mm² almost-square die. As I couldn't find the specifics online, let's say it's 23x26.4mm (that's 607.2mm², close enough, but obviously not accurate). Let's plug that into Caly Technologies' die-per-wafer calculator (sadly only on the
Wayback machine these days). On a 300mm wafer, assuming TSMC's long-reported 0.09 defect rate (which should be roughly applicable for N4, as N4 is a variant of N5, and N5 is said to match N7 defect rates, which were 0.09 several years ago), that results in 87 total dice per wafer, of which ~35 would have defects, and 52 would be defect-free. Given how GPUs are massive arrays of identical hardware, it's likely that all dice with defects are usable in a cut-down form. Let's then assume that half of defect-free dice meet the binning requirements for a fully enabled SKU. That would leave Nvidia with three choices:
- Launch a cut-down flagship consumer SKU at a binning and active block level that lets them use all chips that don't meet binning criteria for a fully enabled chip, and sell all fully enabled chips in higher margin markets (enterprise/workstation etc.) - but also launch a fully enabled consumer SKU later
- Launch a fully enabled consumer SKU
and a cut-down SKU at the same time, with the fully enabled SKU being somewhat limited in quantity and taking some supply away from the aforementioned higher margin markets
- Only ever launch a cut-down consumer SKU, leaving fully enabled chips only to other markets
Nvidia consistently picks the first option among these - the option that goes
hard for maximizing profits above all else, while also necessarily including the iffy move of promising "this is the flagship" just to supersede it 6-12 months later. And that? That's a shitty move, IMO. Is it
horrible? Of course not. But it's explicitly exploitative and cash-grabby at the expense of customers, which makes it shitty.
Still, the volume they need/want to ship on these products, they need to accept a small cut down to produce enough acceptable ones in the given timeframe, save the golden chips for the professional RTX products with more VRAM like the ampere A series products.
That depends on die size and actual yields. As I showed above, with published yields for the process nodes used here, there are still
lots of chips that would meet the criteria for fully enabled SKUs.
Different how? It's worse value than all but 4 available GPUs instead of 2?
Also remember that that chart for some reason only assumes MSRP
or lower rather than the expected and actual reality of prices being MSRP
or higher.
