Nvidia GPU Wafer Allocations *spawn*

[Bondrewd]

We're dealing with bleeding edge tech companies here

Bleeding edge as in foundry stuff.

That why you plan, and do not scramble after shit has hit the fan.

The what.
You only have so much time and manpower to make your shrinks actually be worthwhile.
You don't even consider risky shit a-la 7LPP here.

Do you think they planned for all of this in their A-scenario?

Yeah
G6X is not even a JEDEC thing made for nV duties exclusively.
Those things were always ovens.
What changes was product positioning and die chops, aka the things that change all the time.

 

[CarstenS]

Bleeding edge as in foundry stuff.

You don't say.

The what.

The plan. P-L-A-N.

You only have so much time and manpower to make your shrinks actually be worthwhile.
You don't even consider risky shit a-la 7LPP here.

What shrinks bro? Sidetracking and strawmanning again?

Yeah
G6X is not even a JEDEC thing made for nV duties exclusively.
Those things were always ovens.
What changes was product positioning and die chops, aka the things that change all the time.

Yeah, because everyone and Nvidia especially likes furnaces. Because they shed a good light on your engineering and sell like... hotcakes.

Ok, bro, let's see if I got this right, let me just quote your arguments for reference:

 

[Bondrewd]

At best Nvidia could’ve sold it as a 3080 and asked for $800 and that would’ve been received very poorly

Yeah but 104 chop would still sell gangbusters.

Navi 21 likely had very little to do with their product positioning

Oh it did a lot.
In general we had a lot of SKU shuffling around.

 

[Deleted member 13524]

Also, the 3070 Ti is only 10% faster than a 2080 Ti. At best Nvidia could’ve sold it as a 3080 and asked for $800 and that would’ve been received very poorly. Navi 21 likely had very little to do with their product positioning. Nvidia had to dip into GA102 anyway to comfortably beat Turing.

The 2080 Ti is a TU102 GPU that started at $1000. A fair comparison should be the 2080 non-ti here.

It's possible that Nvidia planned for the GA104 to succeed TU104, GA102 to succeed TU102, i.e., having the full GA104 (3070 Ti) to succeed full TU104 (2080 Super) as RTX 3080 but at a respectable ~35% faster;
With the RX 6800 XT and 6900 XT, they may have had to promote the 3080 to a GA102, and at the same time demote the Ampere Titan RTX to the 3090.

There are some big signs pointing to the 3090 being planned as a Titan card: lots of VRAM, ridiculously large and expensive cooler, over-engineered PCB, power consumption through the roof, etc.

I doubt Nvidia planned from the beginning to sell the 24GB G6X 3090 for just $1500 when they had been selling the GA102 Titan RTX with 24GB G6 for $2500 the year before. That should be reason enough to suspect they did some reshuffling at the high-end.
And the only thing that changed between 2018 and 2020 was that this time AMD had a competitor at that performance range, in the form of Navi 21.

Lets track down TT's credited source:

Why would you fail to quote my last post that refers to the original source from a taiwanese financial news website, which was posted a whole 6 hours before yours?

 

[trinibwoy]

It's possible that Nvidia planned for the GA104 to succeed TU104, GA102 to succeed TU102, i.e., having the full GA104 (3070 Ti) to succeed full TU104 (2080 Super) as RTX 3080 but at a respectable ~35% faster;

It’s extremely unlikely that they planned to ship perfect GA104 dies in volume at series launch. A theoretical GA104 based 3080 would’ve sold well enough but would have been received far less positively given the meagre improvement over the 2080 Ti.

The actual 3080 is a 30% improvement over the 2080 Ti which makes it a lot more attractive proposition. Ampere was competing with Turing far more than Navi.

The SKU scramble post initial launch was likely a response to Navi 21 but those were rounding error moves and didn’t change the overall stack much at all.

 

[Bondrewd]

It’s extremely unlikely that they planned to ship perfect GA104 dies in volume at series launch

Tbh 8LPP yields are like super nice.

Ampere was competing with Turing far more than Navi.

It was competing with Pascal/Maxwell grandpas clinging to their good things.

 

[Deleted member 13524]

A theoretical GA104 based 3080 would’ve sold well enough but would have been received far less positively given the meagre improvement over the 2080 Ti.

The actual 3080 is a 30% improvement over the 2080 Ti which makes it a lot more attractive proposition. Ampere was competing with Turing far more than Navi.

The SKU scramble post initial launch was likely a response to Navi 21 but those were rounding error moves and didn’t change the overall stack much at all.

The RTX 2080 was a 9% boost over the 1080Ti.
The GTX 980 was a 5% boost over the 780 Ti.

A GA104 3080 would be completely fine if it brought a 10% boost over the 2080 Ti.

It’s extremely unlikely that they planned to ship perfect GA104 dies in volume at series launch.

Doesn't need to be perfect. Take away a couple of SMs like they did for RTX 2080 TU104 and the performance difference would be negligible.
By "full" I meant not taking away >8 SMs and some PHYs or slower memory that brings performance down by more than a notch (e.g. 1080 > 1070).

EDIT: a word.