G80 De-Mystified; GeForce 8800GTX/GT

[Pabster]

Bah, I don't care if it takes 300 watts under load. Though I personally do hope they go with an external power "brick" if they really do reach 250W or higher, as most people aren't going to have the necessary high-end power supply to handle that.

High-end graphics has never been a steward of power consumption and I don't look for that to change.

 

[coldpower27]

Originally posted by: Cookie Monster

Originally posted by: Nightmare225

Originally posted by: Cookie Monster
A 7900Gt consumes about 40W~. Funny how the next DX10 cards will consume up to ~250W all by it self. Thats like 6 7900GTs in one.

Ehh? Are these idle readings?

Thats load. From Xbitlabs, the 7900GT consumes 40~W.

Your rounding down....

http://www.xbitlabs.com/articles/video/display/powercolor-x1900gt_5.html

It's actually close to 50W's still excellent compared to the X1900 XTX's value of 120W or the 7900 GTX 84W.

 

[coldpower27]

Originally posted by: Acanthus
And to the people who say the transistor count isnt possible.

TSMCs 65nm ready for mainstream.

"NexsysSM 65-nm process many advantages over other technologies. The 65-nm process is double the gate density and more than 50-percent faster than NexsysSM 90nm. And, because there's minimal process modification there's an easy migration from 90nm."

Yes, but as Nvidia found out with the 0.13 micron process, having a high end product on a untested process is a riskly maneuver, without even smaller less complex chips testing the process out.

At best were hoping for 80nm process on the g80, but currentl the rumors point to a 90nm GPU.

 

[Acanthus]

Originally posted by: coldpower27

Originally posted by: Acanthus
And to the people who say the transistor count isnt possible.

TSMCs 65nm ready for mainstream.

"NexsysSM 65-nm process many advantages over other technologies. The 65-nm process is double the gate density and more than 50-percent faster than NexsysSM 90nm. And, because there's minimal process modification there's an easy migration from 90nm."

Yes, but as Nvidia found out with the 0.13 micron process, having a high end product on a untested process is a riskly maneuver, without even smaller less complex chips testing the process out.

At best were hoping for 80nm process on the g80, but currentl the rumors point to a 90nm GPU.

you simply wont see a GPU with that many transistors at the price target on 80nm, youre talking 3% yields.

 

[coldpower27]

Originally posted by: Acanthus

Originally posted by: coldpower27

Originally posted by: Acanthus
And to the people who say the transistor count isnt possible.

TSMCs 65nm ready for mainstream.

"NexsysSM 65-nm process many advantages over other technologies. The 65-nm process is double the gate density and more than 50-percent faster than NexsysSM 90nm. And, because there's minimal process modification there's an easy migration from 90nm."

Yes, but as Nvidia found out with the 0.13 micron process, having a high end product on a untested process is a riskly maneuver, without even smaller less complex chips testing the process out.

At best were hoping for 80nm process on the g80, but currentl the rumors point to a 90nm GPU.

you simply wont see a GPU with that many transistors at the price target on 80nm, youre talking 3% yields.

Depends, with G70 being 333mm2 or so, an 80nm G80 with an estimated 700 Million Transistors would be ~ 400mm2 or so using G71 Transistor density as a base. Bigger but in line with the increases we have been seeing over time. I would say 3% is a very low estimate. The G70 did drop in price over time despite having a pretty large die, down to 270 ish or so if I recall for the 7800 GT. ATI's products based on R580 and it's large die have also come down in price.

We just don't know what kind of yield rates Nvidia will be getting on a 400mm2 die, as well you have to remember from the looks of the specs, Nvidia doesn't neeed completely fully functional dies. 7/8 or 6/8 for 8800 GTX and 8800 GT respectively.

Also a larger die is inline with a larger memory bus, unless you wish to really have thick layer PCB. With the minimum for a 256Bit Interface being 200mm2 or so, 400mm2 would allow that increase comfortably.

I seriously doubt any flagship product will be 65nm till at least Mid 2007 or later. Mainstream SKU's could be 65nm by Early 2007.

 

[zephyrprime]

Originally posted by: coldpower27With the minimum for a 256Bit Interface being 200mm2 or so, 400mm2 would represents an increase ~ 50% for the side area, enough to allow the increase to 384Bit.

Side area? They use bga connections you know.

 

[coldpower27]

Originally posted by: zephyrprime

Originally posted by: coldpower27With the minimum for a 256Bit Interface being 200mm2 or so, 400mm2 would represents an increase ~ 50% for the side area, enough to allow the increase to 384Bit.

Side area? They use bga connections you know.

Ah ok, yeah after checking side area doesn't make sense. Fixed.

 

[Wreckage]

Originally posted by: coldpower27

I seriously doubt any flagship product will be 65nm till at least Mid 2007 or later. Mainstream SKU's could be 65nm by Early 2007.

8900GT FTW

 

[bobsmith1492]

Originally posted by: coldpower27

Originally posted by: zephyrprime

Originally posted by: coldpower27With the minimum for a 256Bit Interface being 200mm2 or so, 400mm2 would represents an increase ~ 50% for the side area, enough to allow the increase to 384Bit.

Side area? They use bga connections you know.

Ah ok, yeah after checking side area doesn't make sense. Fixed.

Yeah, besides, the die is waaaay smaller than the case; the case is just big enough to fit all the pins on it, and wires are brought out from the die to the pins (on the bottom or sides).

 

[Acanthus]

Originally posted by: bobsmith1492

Originally posted by: coldpower27

Originally posted by: zephyrprime

Originally posted by: coldpower27With the minimum for a 256Bit Interface being 200mm2 or so, 400mm2 would represents an increase ~ 50% for the side area, enough to allow the increase to 384Bit.

Side area? They use bga connections you know.

Ah ok, yeah after checking side area doesn't make sense. Fixed.

Yeah, besides, the die is waaaay smaller than the case; the case is just big enough to fit all the pins on it, and wires are brought out from the die to the pins (on the bottom or sides).

What you said makes no sense to me at all. (CSEE student).

Unless youre talking about the interconnects between the die and the packaging maybe?

 

[Acanthus]

G71 = 271,000,000 transistors, 90nm process, 196 sq mm (14mm x 14mm)

Rumored G80 = 700,000,000 transistors, 90nm process, ~ 494 sq mm (~22.23mm x 22.23mm)

Rumored G80 = 700,000,000 transistors, 80nm process, ~ 308 sq mm (~17.56mm x 17.56mm)

Rumored G80 = 700,000,000 transistors, 65nm process, ~ 243 sq mm (~15.56mm x 15.56mm)

Intel Core 2 Duo = 291,000,000 transistors, 65nm process, 142 sq mm (~11.92mm x 11.92mm)

AMD Athlon X-2 = 154,000,000 transistors, 90nm process, 142 sq mm (~11,92mm x 11.92mm)

ATi Radeon X1950XTX = 384,000,000 transistors, 90nm process, and im not doing any more math, ive proven my point already.

In other words, its a huge 80nm chip, and an insurmountably huge 90nm chip.

Without even considering power, clockspeed, and heat, 90nm just looks bad to begin with.

 

[Smartazz]

Originally posted by: Acanthus
G71 = 271,000,000 transistors, 90nm process, 196 sq mm (14mm x 14mm)

Rumored G80 = 700,000,000 transistors, 90nm process, ~ 494 sq mm (~22.23mm x 22.23mm)

Rumored G80 = 700,000,000 transistors, 80nm process, ~ 308 sq mm (~17.56mm x 17.56mm)

Rumored G80 = 700,000,000 transistors, 65nm process, ~ 243 sq mm (~15.56mm x 15.56mm)

Intel Core 2 Duo = 291,000,000 transistors, 65nm process, 142 sq mm (~11.92mm x 11.92mm)

AMD Athlon X-2 = 154,000,000 transistors, 90nm process, 142 sq mm (~11,92mm x 11.92mm)

ATi Radeon X1950XTX = 384,000,000 transistors, 90nm process, and im not doing any more math, ive proven my point already.

In other words, its a huge 80nm chip, and an insurmountably huge 90nm chip.

Without even considering power, clockspeed, and heat, 90nm just looks bad to begin with.

What if they go with 2 90nm chips on the same board like they did with the 7950GX2. Is this still a possibility?

 

[Nightmare225]

Originally posted by: Smartazz

Originally posted by: Acanthus
G71 = 271,000,000 transistors, 90nm process, 196 sq mm (14mm x 14mm)

Rumored G80 = 700,000,000 transistors, 90nm process, ~ 494 sq mm (~22.23mm x 22.23mm)

Rumored G80 = 700,000,000 transistors, 80nm process, ~ 308 sq mm (~17.56mm x 17.56mm)

Rumored G80 = 700,000,000 transistors, 65nm process, ~ 243 sq mm (~15.56mm x 15.56mm)

Intel Core 2 Duo = 291,000,000 transistors, 65nm process, 142 sq mm (~11.92mm x 11.92mm)

AMD Athlon X-2 = 154,000,000 transistors, 90nm process, 142 sq mm (~11,92mm x 11.92mm)

ATi Radeon X1950XTX = 384,000,000 transistors, 90nm process, and im not doing any more math, ive proven my point already.

In other words, its a huge 80nm chip, and an insurmountably huge 90nm chip.

Without even considering power, clockspeed, and heat, 90nm just looks bad to begin with.

What if they go with 2 90nm chips on the same board like they did with the 7950GX2. Is this still a possibility?

Always a possiblity, I am hoping they don't put that experience to waste, it worked very well for them this gen...

 

[Acanthus]

Originally posted by: Smartazz

Originally posted by: Acanthus
G71 = 271,000,000 transistors, 90nm process, 196 sq mm (14mm x 14mm)

Rumored G80 = 700,000,000 transistors, 90nm process, ~ 494 sq mm (~22.23mm x 22.23mm)

Rumored G80 = 700,000,000 transistors, 80nm process, ~ 370 sq mm (~19.24mm x 19.24mm)

Rumored G80 = 700,000,000 transistors, 65nm process, ~ 243 sq mm (~15.56mm x 15.56mm)

Intel Core 2 Duo = 291,000,000 transistors, 65nm process, 142 sq mm (~11.92mm x 11.92mm)

AMD Athlon X-2 = 154,000,000 transistors, 90nm process, 142 sq mm (~11,92mm x 11.92mm)

ATi Radeon X1950XTX = 384,000,000 transistors, 90nm process, and im not doing any more math, ive proven my point already.

In other words, its a huge 80nm chip, and an insurmountably huge 90nm chip.

Without even considering power, clockspeed, and heat, 90nm just looks bad to begin with.

What if they go with 2 90nm chips on the same board like they did with the 7950GX2. Is this still a possibility?

It very well could be done that way, but then the 384bit memory bus and other rumors like clockspeed dont line up.

Edit: 80nm math corrected, because apparently it was relevant.

 

[Smartazz]

Originally posted by: Acanthus

Originally posted by: Smartazz

Originally posted by: Acanthus
G71 = 271,000,000 transistors, 90nm process, 196 sq mm (14mm x 14mm)

Rumored G80 = 700,000,000 transistors, 90nm process, ~ 494 sq mm (~22.23mm x 22.23mm)

Rumored G80 = 700,000,000 transistors, 80nm process, ~ 308 sq mm (~17.56mm x 17.56mm)

Rumored G80 = 700,000,000 transistors, 65nm process, ~ 243 sq mm (~15.56mm x 15.56mm)

Intel Core 2 Duo = 291,000,000 transistors, 65nm process, 142 sq mm (~11.92mm x 11.92mm)

AMD Athlon X-2 = 154,000,000 transistors, 90nm process, 142 sq mm (~11,92mm x 11.92mm)

ATi Radeon X1950XTX = 384,000,000 transistors, 90nm process, and im not doing any more math, ive proven my point already.

In other words, its a huge 80nm chip, and an insurmountably huge 90nm chip.

Without even considering power, clockspeed, and heat, 90nm just looks bad to begin with.

What if they go with 2 90nm chips on the same board like they did with the 7950GX2. Is this still a possibility?

It very well could be done that way, but then the 384bit memory bus and other rumors like clockspeed dont line up.

When it's dual cards doesn't nvidia report it as 512bit instead of 256bit. Maybe it's the same approach. Who knows at this point?

 

[Acanthus]

Originally posted by: Smartazz

Originally posted by: Acanthus

Originally posted by: Smartazz

Originally posted by: Acanthus
G71 = 271,000,000 transistors, 90nm process, 196 sq mm (14mm x 14mm)

Rumored G80 = 700,000,000 transistors, 90nm process, ~ 494 sq mm (~22.23mm x 22.23mm)

Rumored G80 = 700,000,000 transistors, 80nm process, ~ 308 sq mm (~17.56mm x 17.56mm)

Rumored G80 = 700,000,000 transistors, 65nm process, ~ 243 sq mm (~15.56mm x 15.56mm)

Intel Core 2 Duo = 291,000,000 transistors, 65nm process, 142 sq mm (~11.92mm x 11.92mm)

AMD Athlon X-2 = 154,000,000 transistors, 90nm process, 142 sq mm (~11,92mm x 11.92mm)

ATi Radeon X1950XTX = 384,000,000 transistors, 90nm process, and im not doing any more math, ive proven my point already.

In other words, its a huge 80nm chip, and an insurmountably huge 90nm chip.

Without even considering power, clockspeed, and heat, 90nm just looks bad to begin with.

What if they go with 2 90nm chips on the same board like they did with the 7950GX2. Is this still a possibility?

It very well could be done that way, but then the 384bit memory bus and other rumors like clockspeed dont line up.

When it's dual cards doesn't nvidia report it as 512bit instead of 256bit. Maybe it's the same approach. Who knows at this point?

That would mean that they would actually reduce the number of memory controllers per core, which again doesnt make sense.

And you simply wont get 1.5ghz on a 500mm die.

 

[josh6079]

What if they go with 2 90nm chips on the same board like they did with the 7950GX2. Is this still a possibility?

Perhaps, but the 7950GX2 doesn't have 2 90 nm chips on the same board. It would be more like Asus's dual-core 7800GTX in that regard.

 

[bluemax]

Insane.... I'm just getting used to seeing the 7x00-series around and now there's all new stuff? Sheesh! Moving fast.....


... I'd love to see more energy-efficient designs though. My new PC isn't going to have a hefty power supply - then again, I can't afford top-of-the-line graphics anyways. Affordable, mid-range stuff with low power consumption and fanless designs are my favorite.

 

[Smartazz]

Originally posted by: josh6079

What if they go with 2 90nm chips on the same board like they did with the 7950GX2. Is this still a possibility?

Perhaps, but the 7950GX2 doesn't have 2 90 nm chips on the same board. It would be more like Asus's dual-core 7800GTX in that regard.

Yeah, I saw that Asus put dual G70's on the same board. The Extreme 7800GT, something like that.

 

[coldpower27]

Originally posted by: Acanthus
G71 = 271,000,000 transistors, 90nm process, 196 sq mm (14mm x 14mm)

Rumored G80 = 700,000,000 transistors, 90nm process, ~ 494 sq mm (~22.23mm x 22.23mm)

Rumored G80 = 700,000,000 transistors, 80nm process, ~ 308 sq mm (~17.56mm x 17.56mm)

Rumored G80 = 700,000,000 transistors, 65nm process, ~ 243 sq mm (~15.56mm x 15.56mm)

Intel Core 2 Duo = 291,000,000 transistors, 65nm process, 142 sq mm (~11.92mm x 11.92mm)

AMD Athlon X-2 = 154,000,000 transistors, 90nm process, 142 sq mm (~11,92mm x 11.92mm)

ATi Radeon X1950XTX = 384,000,000 transistors, 90nm process, and im not doing any more math, ive proven my point already.

In other words, its a huge 80nm chip, and an insurmountably huge 90nm chip.

Without even considering power, clockspeed, and heat, 90nm just looks bad to begin with.

Your 80nm process figure is off.

Assuming perfect optical shrinking and the exact same density for the 90nm process.

G71 = 278 Million Transistors on 196mm2

Transistor Density is 1,418,367 per mm2.

Die Size on 90nm = 494mm2

Die Size on 80nm = 6400/8100 = 79% of the size of 90nm die if assuming optical shrinkage is perfect.

Die Size on 80nm = 390mm2

Die Size on 65nm = 4225/6400 66% of the 80nm node = 257mm2.

You haven't proven your point at all, CPU's are a different category all together, not to mention your quoting the Manchester Die Size, which is outdated AMD is using Rev F. processors which have 153.8 Million Transistors on a die size of 183mm2. In CPU's the same core that is used in high end products are used in low end.

For example Pentium D 805 goes for 93US MSRP but is a die size of 206mm2.

For sure it is not a 65nm product, at best we can hope for is a 80nm product at this time and that is wrong from the current information we have, while a near 500mm2die size product is large, it's fine for the price points of 449US and 649US of the 8800 Series.

It's looks bad for sure, but to add the performance to 2x as well as increase the functionality Nvidia and ATi have no choice.

Each generation going forward has been larger and larger dies from the Geforce 4 Ti Series and Onward,

Here is something interesting:

http://xbitlabs.com/news/video/display/20060920234449.html

It seems high end product bring in the majority of the revenue levels, despite their small marketshare in numbers.

If the data from Xbitlabs is correct, it wouldn't be bad to make die that are 4x as big as the mainstream for example since they bring in 74% of the market revenue, with mainstream SKU's commanding 8% so over 9x more then mainstream SKU's like the 7600GT/X1600 XT.

 

[Cookie Monster]

Im guessing that G80 will be on 80nm.

nVIDIA always tested new process on a low end chip. Hence before G70 --> G71, they tested 90nm using the 6150/6100.
Now this time, with rumours of being released around nov/dec, im guessing they are moving to 80nm as they have already released the first 80nm products. (7700 go, and 7650 series to come).

 

[Acanthus]

Originally posted by: coldpower27

Originally posted by: Acanthus
G71 = 271,000,000 transistors, 90nm process, 196 sq mm (14mm x 14mm)

Rumored G80 = 700,000,000 transistors, 90nm process, ~ 494 sq mm (~22.23mm x 22.23mm)

Rumored G80 = 700,000,000 transistors, 80nm process, ~ 308 sq mm (~17.56mm x 17.56mm)

Rumored G80 = 700,000,000 transistors, 65nm process, ~ 243 sq mm (~15.56mm x 15.56mm)

Intel Core 2 Duo = 291,000,000 transistors, 65nm process, 142 sq mm (~11.92mm x 11.92mm)

AMD Athlon X-2 = 154,000,000 transistors, 90nm process, 142 sq mm (~11,92mm x 11.92mm)

ATi Radeon X1950XTX = 384,000,000 transistors, 90nm process, and im not doing any more math, ive proven my point already.

In other words, its a huge 80nm chip, and an insurmountably huge 90nm chip.

Without even considering power, clockspeed, and heat, 90nm just looks bad to begin with.

Your 80nm process figure is off.

Assuming perfect optical shrinking and the exact same density for the 90nm process.

G71 = 278 Million Transistors on 196mm2

Transistor Density is 1,418,367 per mm2.

Die Size on 90nm = 494mm2

Die Size on 80nm = 6400/8100 = 79% of the size of 90nm die if assuming optical shrinkage is perfect.

Die Size on 80nm = 390mm2

Die Size on 65nm = 4225/6400 66% of the 80nm node = 257mm2.

You haven't proven your point at all, CPU's are a different category all together, not to mention your quoting the Manchester Die Size, which is outdated AMD is using Rev F. processors which have 153.8 Million Transistors on a die size of 183mm2. In CPU's the same core that is used in high end products are used in low end.

For example Pentium D 805 goes for 93US MSRP but is a die size of 206mm2.

For sure it is not a 65nm product, at best we can hope for is a 80nm product at this time and that is wrong from the current information we have, while a near 500mm2die size product is large, it's fine for the price points of 449US and 649US of the 8800 Series.

It's looks bad for sure, but to add the performance to 2x as well as increase the functionality Nvidia and ATi have no choice.

Each generation going forward has been larger and larger dies from the Geforce 4 Ti Series and Onward,

Here is something interesting:

http://xbitlabs.com/news/video/display/20060920234449.html

It seems high end product bring in the majority of the revenue levels, despite their small marketshare in numbers.

If the data from Xbitlabs is correct, it wouldn't be bad to make die that are 4x as big as the mainstream for example since they bring in 74% of the market revenue, with mainstream SKU's commanding 8% so over 9x more then mainstream SKU's like the 7600GT/X1600 XT.

Uhh, you cant just use %ages to get the answer...

Algebra 101.

 

[coldpower27]

Originally posted by: Acanthus
Uhh, you cant just use %ages to get the answer...

Algebra 101.

Yeah you can, with the way I have done it.

 

[Coldkilla]

Excuse me, but how can someone like me understand What your all talking about?! I've just started PC Science major in college, and i'd like to know first of all what all this means! and second, how to compair it to something else, (real world example) to visualize wtf is talking about! all i know is the number 8800 is higher than the 7800..

 

[Acanthus]

a(sq) + b(sq) =/= (a+b)(sq)