Why should I? I don't need to prove anything to anyone. Consider it just a fair warning and if things turn out as the local gang here expects simply rejoice in a handful of months. In any other case you won't see me chearleading over any silly "told you so's" so pick your poison.
I've the right to voice my opinion in any way it suits me as long as I don't violate the local TOS. I'm not obliged to do anything you tell me and I don't have to tolerate any direct or indirect verbal attacks either.
If you want to talk market share (?), by making cards that sell by the millions. You do realize the figures are based on units sold, not monetary value ?
That's why AMDs market share only improved slightly last year, despite the introduction of Fury-X, Fury, the Nano and 8GB 390X. Sure, they sold ok, but in terms of volume they're a fraction of the overall market. And in terms of sheer numbers, completely drowned out by the 750/950/960/970s of this world.
If AMD want to regain market share, it's those cards they need to compete with. Not the 980ti. If they introduced another $500 'ultra enthusiast' card I doubt it would change their market share by more than a few percent, even if it was the biggest/fastest card ever.
To significantly change their market share they need to sell millions of cards, not tens of thousands. But 'enthusiasts' never seem to understand that. Go figure.
To find the right answers you must ask the right questions.
Let's just say I hear a couple of things for A or B from time to time because I'm an old dog around for decades for anything 3D. That doesn't mean that I don't make mistakes or interpret things the wrong way from time to time, but I don't really make up random stuff either.
That said back to the topic: in all likeliness neither Greenland nor GP100 will make it to desktops this year because of too high manufacturing costs. For which the first shouldn't be as large as the latter either and that is not even as big as the current GM200.
In order for a FF process based perfomance chip to make an SKU with a bit of a profit at a =/<$350 MSRP the chip should not cost more than say 80 to manufacture. If you want to believe it's possible with any FinFET process from the kickstart be my guest.
except that $350 isn't what AMD sells it for.
take 40-50 for the retailer, 40-50 in profit for the AIB. Than take the remaining $250-270 or so and subtract the cost of everything besides the GPU. Than you have the cost AMD can sell the GPU for.
At $330 for the 390 it's probably breaking down to:
40 retailer profit
40 AIB profit
~80 dollars to produce.
So you have about 170 dollars for all the components except the GPU itself and any margin AMD is making.
I'm guessing AMD ends up making about $80-100 per 390 die sold.
This 232mm^2 chip probably costs about 25-30% more per chip than Hawaii assuming equal yields. Which ends up leaving about $125-170 for everything but the AIB/retail markups and the cost to produce the GPU. What margin is there? Probably closer to the $125 mark. At which point AMD probably makes around 30-50 dollars per die.
-Q
I presume that is a FEOL issue however? The poor hot transistors got sweaty in close proximity and there was nowhere to vent off. 14/16nm FF fixes that.
And the solution was to put bigger gaps between transistors, hence the 28-20 density improvement wasn't great, articles mentioned this in vague terms.
I agree. The A8 had 2 billion in 89mm^2 on 20nm, I think they announced that one. Of course, that's to 1 significant digit :/ Ditto the A8X is 3 billion in 128mm^2 on 20nm.I don't know how he reached those supposed 3.6b but he's most likely wrong.
Even ignoring everything, even 20->16nm, 147 is 1.15x the area, so 3.45B transistors. That's a pretty safe minimum bound IMO. TBH I've forgotten what we are talking about now.
Fair point, although it may still suggest the ballpark target floorplan area AMD wanted, say 210 - 250.Again for those suppposed 232mm2, when I asked the answer was that it comes from floorplans that never materialized.
Unless AMD were forced to go lower because of yield issues above 200mm^2 - that's another snake to add to the snake pit...
This is probably the sane solution to the aboveI personally ALWAYS have modest expectations in order to always get pleasantly surprised![]()
![]()
ROFLquite similar here during the course of the debate. Truth be told there are a ton of unknown factors. Even if we'd know what each FF wafer costs for an indicative value, it's still fairly useless unless you know exact yields. However when a process is as young at 14LPP and/or 16FF+ it's pretty much a safe bet to expect worse yields than on very mature processes like 28HP TSMC. Folks that expect f.e. today a 400mm2 die at 16FF+ to be cheaper to manufacture than a 600mm2 die at 28HP might want to shake their heads.