Nvidia's Next Gen Speculation

[JohnJohn]

Well since all of the news highlights are on HD6000 & Sandy bridge & Fusion and yadayadayada, i think it is time to speculate about Nvidia's Next big thing.

A couple of days ago we were notified that The possible dual-gpu solution from nvidia has been cancelled.

Sorry for ressurecting buried thread, but something new is about GTX 490. We already have complete card design, manufactured, tested, packed! Launch time was planned around this weekend, but NVIDIA cancelled this card for ever!

So the only info we have to speculate on is this:

We also know that NVIDIA is currently planning to GF110-koodinimellistä top GPUs, designed in line with that of consumer-class GeForce GTX 460's GF104 GPUs, namely Streaming Multiprocessor Modules edified 48 CUDA kernel. Teoriassa 10, 12 tai 16 SM-moduulilla CUDA-ytimien lukumäärä olisi 480, 576 tai 768. Theoretically, 10, 12 or 16 SM module CUDA-cores would be 480, 576 or 768 CUDA-ytimien ja ROP-yksiköiden välinen suhde on 12:1. CUDA kernels and BOB-staff ratio is 12:1.

http://plaza.fi/muropaketti/nvidian-...-suunnitteilla
http://translate.google.com/translat...-suunnitteilla

A supposed new chip called "Gf110" made from the Gf104 architecture with up to 768 Cuda cores. I now welcome our new 700mm2+ overlords!

Update 09/21/10
"Exciting stuff at the end of Jen-Hsun Huang's GTC keynote as Nvidia's CEO has presented a roadmap depicting the next two GPU architectures currently in the works at the Santa Clara company. Fermi's successors are known as Kepler and Maxwell and are scheduled to launch in 2011 and 2013, respectively.
Kepler is confirmed to be made for the 28nm process, will enter production next year, and is supposed to have 3 to 4 times the performance/watt of Fermi. Kepler will also implement various technologies to limit CPU bottlenecks but those will be detailed at a later date (maybe only upon release)."
http://www.tcmagazine.com/tcm/news/h...ming-next-year

[Manicdan]

due to the massive chip sizes nvidia uses, i petition that we change the standard of measurement to be in meters squared, so its .0007m2

[Iconyu]

I don't think TSMC would allow nvidia to attempt a 700mm sq part, but it would be fun to see.

[kaktus1907]

700mm2 chip means bye bye nvidia..

[dengyong]

I'd like to have better performance with much less power draw and heat.

[ElSel10]

Yeah, I'm sure Nvidia is planning a 700mm^2+ GPU. Get real guys.

[-Sweeper_]

768 SPs it too much for 40nm, but 576 SPs (@ GF104 scheme) sounds doable..

if they get those 20% more SPs over GTX480 working at 700mhz+ under 300w TDP, that could do well against cayman XT...

[Ejizz]

Thought I saw a pig go past the bedroom window then...

[Nintendork]

I saw nvidia wanting to open an OEM project:

They're planning to put their stickers in some product as a rebrand:

Nvidia HD7800gtx 28nm.

[hurleybird]

Nvidia needs to pull an RV770.

[Jowy Atreides]

Yeah, I'm sure Nvidia is planning a 700mm^2+ GPU. Get real guys.

Why not?

8800


280



480

[Metroid]

Why not?

[tajoh111]

768 SPs it too much for 40nm, but 576 SPs (@ GF104 scheme) sounds doable..

if they get those 20% more SPs over GTX480 working at 700mhz+ under 300w TDP, that could do well against cayman XT...

It depends really on the size of how big GF104 is right now. If we take the 366mm figure as true then the specs of this rumor are out on of the question. But if GPUz reporting of 322mm is true than it becomes more possible. Gf106 is 222mm2. and it has half the shaders, half the texture units, 8 less ROP, and a simpler memory controller. NV is able to add 192, shaders, 32 texture units, 8 more ROPS for 100mm more(the full gf104). Efficiency of adding things seems to going up with gf 104,106, 108 derivatives. The gf108 being the least efficient and the gf104 being the most efficient. For 200mm in extra budget, if we take the gf104 vs gf106 as an estimate, they might be able to add 384 more shaders, 16 or more ROP and double the texture units.

Power consumption might be another story, but the power consumption difference might be tolerable if they keep clocks below 700.

I sense another strip down part even if this is true. Forget about 768, I can see at best them getting 720 shaders working with decent yield.

[Ejizz]

^^^
I thought GF104 was more like 380'ish mm2 and is larger than Cypress?

Edit:

I sense another strip down part even if this is true. Forget about 768, I can see at best them getting 720 shaders working with decent yield.

Please, enough of the fooling around and joking, can we be please be serious now?

[tajoh111]

^^^
I thought GF104 was more like 380'ish mm2 and is larger than Cypress?

Thats the very worst with Charlie said before the chip came out. He also said it would consume 210 watts which is completely off.

http://www.semiaccurate.com/2010/06/...106-and-gf108/

Even he conceded 367mm

http://www.semiaccurate.com/2010/07/...-has-huge-die/

If you look at a GPU-z screen it says 332mm2.

With better planning and refining the process they should be able to get smaller as the chip itself has 10% less transisters than cypress.

[G.Foyle]

Guys, this is XS - why don't you go and remove the IHS to end the confusion about die size? I would do it if I had a card.

[570091D]

here we go again... with nvidia, you never know until they actually start selling the product. there are so many crazy fanboi and troll rumors flying around it's almost impossible to cut through the s#!t.

having said that, it would make sense for them to lay gf100 to rest, it served it's purpose (learing experiance). gf104 has the perfromance people are looking for without the high power consumption and it's cheap to make. why not build on a winning foundation?

[Olivon]

GTC starts today, maybe JHH will tell something

[K404]

We've still to see the 512SP part. I *think* nV are sailing too close to the wind with the current top-end silicon. They need a respin/ silicon revision to get things truely in check. Even at that, I dont think they're overflowing with options.

[spursindonesia]

768 SPs it too much for 40nm, but 576 SPs (@ GF104 scheme) sounds doable..

if they get those 20% more SPs over GTX480 working at 700mhz+ under 300w TDP, that could do well against cayman XT...

Agreed, considering that a 192 SP 24 ROP part GF 106 is already 240 mm^2 in diesize, only fanbois would believe that GF 104 is just 322 mm^2 part, while 367 mm^2 of diesize is MUCH more logical and believable.

OTOH, a 576 48 ROP 384 bit part developed from GF 104 basis is also quite possible even using 367 mm^2 as diesize basis, i think the added diesize would be in the neighborhood of 150-160 mm^2. With the added time and experience, pushing that big diesize out of TSMC 40 nm is still doable IMHO, and while the performance might only just been able to match Cayman @380 mm^2, the WOW and pride factor must be considered for them so they won't lose anymore fanbois and lovers around the globe. Remember, they have once designed a 576 mm^2 monster of GT 200, so it has happened in the past and can verywell happen yet again in the future.

[Baron_Davis]

I just want a small, powerful card that can make laptops almost as good as desktops for gaming...is that too much to ask??

yes...

[Frontl1ne]

I just want a small, powerful card that can make laptops almost as good as desktops for gaming...is that too much to ask??

yes...

Real men use desktops for gaming.

[LordEC911]

Not going to be over 600mm2, GT200 was right at the limit of what is possible.

[Manicdan]

I just want a small, powerful card that can make laptops almost as good as desktops for gaming...is that too much to ask??

yes...

what you want is a laptop that has an external x16 pcie slot on the docking station with its own PSU, so you dock your laptop and it uses whatever video card you want.

is it too much to ask, no
so what hasnt it happened yet?

[turtletrax]

Real men use desktops for gaming.

QFT