Kepler Taped Out

[Copper]

Originally Posted by kac

He probably meant "unfixable" from a market perspective. Whenever you've got die 30% bigger than your competition yet you've got performance that's only -5 to 10% higher you've got a problem.

there ya go.

[Poisoner]

Originally Posted by distinctively

+1 to the analogy. I believe the Rv670 also bumped up to DX10.1 from DX10.



The nVidia internal code name for the GF110 was actually GF100b. The GF110 is really nothing more than a base layer respin of the GF100. Nothing changed except fixing the leaking areas.



I agree that it would be a waste of money if they just completely dumped the Fermi architecture. I don't think they can maintain that arch for long unless the professional side can continue support the gaming side. Without the professional sales, nVidia is in the red for the year AFAIK.

Don't forget that gf100 and gf104 are different themselves.

[andrewuwe]

The really odd thing is that you can still buy gtx480. I know it is only a cut down failed gpu die.

Why and how did they make so many of them?

[TakeANumber]

9000 risk wafers...or slow, slow sales...take you pick

[distinctively]

Originally Posted by andrewuwe

The really odd thing is that you can still buy gtx480. I know it is only a cut down failed gpu die.

Why and how did they make so many of them?

There weren't that many made. The high end never sells in abundance so its easy to have leftovers. Now they are overshadowed by nVidia's newer chips, therefore undesirable. There was also probably stock leftover, intended for the professional side that now is used for video cards instead.

[Nerdmaster]

Originally Posted by distinctively

This sounds like a replay of Fudo making excuses for Fermi. Remember when Fudo claimed that Fermi would launch in November unless they needed extra time to make sure the drivers were solid.

They even told that Ati was in a hurry with 5xxx because they were afraid of a fermi launch. I will try to find the article it came out at summer 2009.

Originally Posted by kac

He probably meant "unfixable" from a market perspective. Whenever you've got die 30% bigger than your competition yet you've got performance that's only -5 to 10% higher you've got a problem.

If you have a die 30% bigger you need more than 30% more performance because it costs more than 30% to produce if you want the same profits. Remember bigger chips = lower yields. That is why fermi needed sth like 40% more performance just to be as good chip as cypress (from market perspective always).

[265586888]

Not only "a hurried launch", but with "non-native DirectX 11 support"...
AMD's reply was "How many DirectX 11 GPUs were shipped from us in 2 years?"

Oh, how I love the satire...

[distinctively]

Originally Posted by Copper

there ya go.

I wish you would say that when you have a large supply of bacon. Wheelbarrow quantities would be good.

[laurent]

Originally Posted by distinctively

There was also probably stock leftover, intended for the professional side that now is used for video cards instead.

That'd be interesting as that'd mean with proper BIOS/drivers you get higher DP flops for GPGPU. Anyone wants to give that a try?

[CRoland]

Originally Posted by laurent

That'd be interesting as that'd mean with proper BIOS/drivers you get higher DP flops for GPGPU. Anyone wants to give that a try?

Wouldn't that in principle be true for all of them?

They use the exact same die, but I don't know if they choose the DP rate in software, firmware or by actual hardware modification.