NVIDIA GF100 & Friends speculation

[Vincent]

Oh, you haven't heard? the chip is going back to TSMC in April for it's first full re-spin!

Really !


Nvidia do need a lineup of the mid-range GPUs for its own sake.

 

[Jawed]

Now, I don't know anything for sure, but I trust someone when he says it's the same size as GT200 over anyone doing damage control on NV's behalf.

Was that person specifically quizzed on it being the same size as GT200, rather than GT200b?

Jawed

 

[trinibwoy]

Oh, you haven't heard? the chip is going back to TSMC after the "launch" for it's first full re-spin!

Hence the no-to-limited availability etc. until June/July. It also is the basis for Charlies "handful" of chips for partners, A3 will be here for launch and B1 will be the "shipping" product.

Isn't that old news? My question to Rys wasn't about the consequences of the screwup, I was asking about the screwup itself.

 

[neliz]

Was that person specifically quizzed on it being the same size as GT200, rather than GT200b?

Jawed

He has been quite adamant over it's size and performance target ever since last summer. Having a spotless track record makes me go "no no no! you're doing it all wrong!" when I see GF100 being estimated a gt200b size.

It could only be ~500mm2 if nvidia somehow managed to match amd's transistor density on 40nm (re: Richard Huddy's statement that GF100 is 50% bigger than cypress; in transistor count that would 334mm2/2,1bT : 507mm2/3,2bT if there was a 1:1 ratio)
It would be my layman's judgment that anything north of 507mm2 would be a "safe bet"

 

[Rys]

What recent delays? There was more stuff going awry after A3?

As far as I can tell, there was originally a plan to launch at CeBIT, rather than just demo there. So a push back of a few weeks, with the net result being the zero-day review coverage should be pretty aggressive in terms of number of publications with hardware. God knows what neliz is on about.

 

[Jawed]

He has been quite adamant over it's size and performance target ever since last summer. Having a spotless track record makes me go "no no no! you're doing it all wrong!" when I see GF100 being estimated a gt200b size.

So you're saying he specifically said GT200, not GT200b. OK.

It could only be ~500mm2 if nvidia somehow managed to match amd's transistor density on 40nm

That's just a red herring. See the recent confusion over defect density as reported by TSMC.

(re: Richard Huddy's statement that GF100 is 50% bigger than cypress; in transistor count that would 334mm2/2,1bT : 507mm2/3,2bT if there was a 1:1 ratio)
It would be my layman's judgment that anything north of 507mm2 would be a "safe bet"

"Info" from AMD should be ignored though.

Anyway, if JHH counts as a source:

http://forum.beyond3d.com/showpost.php?p=1343760&postcount=406

On the size of Fermi, Jensen stated, "It is only big right now, because it is the biggest chip ever built."

Which makes it bigger than GT200

Jawed

 

[Ailuros]

I can't really understand why it's so damn important whether the die is 500 or 550mm2 but that's just me.

 

[rpg.314]

Oh, you haven't heard? the chip is going back to TSMC after the "launch" for it's first full re-spin!

Hence the no-to-limited availability etc. until June/July. It also is the basis for Charlies "handful" of chips for partners, A3 will be here for launch and B1 will be the "shipping" product.

I am assuming that it will take atleast 2 months for the midrange Fermi. If B1 ships in June/July, then it should be Aug/Sept before midrange fermi will come anywhere close to shipping. Almost a complete year behind Cypress/Juniper. More late than R600.

On the up side, with a silicon respin, there is a good possibility of higher clocks and yields.

On the down side, as has been speculated in the Evergreen+1 thread, new AMD parts *may* launch around that time. All in all, lot of catching up to do for nv if this is true.

 

[Rys]

Which makes it bigger than GT200

He could just be going on transistor count

 

[chavvdarrr]

Oh, you haven't heard? the chip is going back to TSMC after the "launch" for it's first full re-spin!

Hence the no-to-limited availability etc. until June/July. It also is the basis for Charlies "handful" of chips for partners, A3 will be here for launch and B1 will be the "shipping" product.

Strange.
Reading this lines my 6th sense sees irony.

 

[Sxotty]

Oh, you haven't heard? the chip is going back to TSMC after the "launch" for it's first full re-spin!

Hence the no-to-limited availability etc. until June/July. It also is the basis for Charlies "handful" of chips for partners, A3 will be here for launch and B1 will be the "shipping" product.

Why in the world would it go back after the launch? Why would it not go back immediately upon deciding they would be doing another revision?

 

[Jawed]

He could just be going on transistor count

Yeah, that was suggested the last time I posted that.

I did an estimate from the GF100 die shot of it being upto around 500mm² (I estimated 480mm² based upon the PCI Express interface, something that should be constant size across many chips):

http://forum.beyond3d.com/showpost.php?p=1344399&postcount=483
http://forum.beyond3d.com/showpost.php?p=1344504&postcount=499

I should point out in retrospect that Fellix's technique, based upon GDDR5 interfaces, is no good because we now know that these interfaces change in size with target clocking.

Jawed

 

[eastmen]

Why in the world would it go back after the launch? Why would it not go back immediately upon deciding they would be doing another revision?

they aren't done fixing the bugs yet ?

 

[trinibwoy]

As far as I can tell, there was originally a plan to launch at CeBIT, rather than just demo there. So a push back of a few weeks, with the net result being the zero-day review coverage should be pretty aggressive in terms of number of publications with hardware. God knows what neliz is on about.

I see. So by aggressive do you mean lots of sites will have hardware in hand to review before launch? If so, bring on the leaks!

 

[Jawed]

Presumably NVidia Focus Group members will also get their cards.

Jawed

 

[Razor1]

Yeah, that was suggested the last time I posted that.

I did an estimate from the GF100 die shot of it being upto around 500mm² (I estimated 480mm² based upon the PCI Express interface, something that should be constant size across many chips):

http://forum.beyond3d.com/showpost.php?p=1344399&postcount=483
http://forum.beyond3d.com/showpost.php?p=1344504&postcount=499

I should point out in retrospect that Fellix's technique, based upon GDDR5 interfaces, is no good because we now know that these interfaces change in size with target clocking.

Jawed

Jensen also stated that it would keep their margins at the same level so I'm more inclined to go with ~480mm2. As I stated the die size wasn't even revealed to AIB when we first started talking about die size and wafer counts and what not.

 

[trinibwoy]

Jensen also stated that it would keep their margins at the same level so I'm more inclined to go with ~480mm2. As I stated the die size wasn't even revealed to AIB when we first started talking about die size and wafer counts and what not.

Given the volumes people are talking about will Fermi based products even matter in Q2 financials?

 

[Malo]

Why in the world would it go back after the launch? Why would it not go back immediately upon deciding they would be doing another revision?

Please investors? Try to make some money from existing silicon? Try to actually convince consumers they have a product so they don't go out and buy an ATI card instead?

 

[Rys]

I see. So by aggressive do you mean lots of sites will have hardware in hand to review before launch? If so, bring on the leaks!

Not sure about lots, but more than usual I think. Looks like they're all on track to get GTX 480 SLI as well.

So the entire working set of chips from thousands of wafers on two or three spins over many months, some chips bigger than others, with hundreds of watts of board power variation, all in the hands of websites and journalists and thus none available to the public whatsoever until July. Or is that June? Personally, I'm going for B2 in August.

 

[GZ007]

Some Cebit pictures http://en.expreview.com/2010/03/02/fermi-based-geforce-gtx470-pictured-at-cebit/6800.html.
The second air intake/output trough the pcb is a great idea even if it wouldnt be a gtx470 .