Leaked GT212 specification

[Cookie Monster]

GT212
40nm process technology
<300 mm^2
<1.8 billion transistors
384 SPs
96 TMUs
32 ROPs
256bit Memory Interface
GDDR5

Release date ~ Q2/09

Some other notes:
-32 Sps per cluster.
-12 Clusters resulting in 96 TMUs
-To retain the 32ROPs, 64bit channel memory controllers per 8 ROPs.
-No ALU changes. No DX10.1
-Smaller die, and maybe shorter PCB length due to the decrease in complexity of the PCB thanks to the 256bit bus.

Source

IMHO, this specification sounds reasonable. Using GDDR5 rated at 1250MHz, it would result in a bandwidth of 160GB/s. Im guessing that its shader domain is clocked at > 1.5GHz effectively making it 60% faster than the GTX285.

 

[Tempered81]

dp

 

[Tempered81]

thought gt212 were mainstream variants of gt200

guess i was wrong. wouldn't be surprised with how many curveballs you get in the gpu game these days.

Good news!

 

[ArchAngel777]

I'll be dissapointed if it is a 256 bit interface. Not that I think the current GTX is memory bandwidth limited, But I do think it would be if the shader power was increased further as these specs might indicate.

 

[SlowSpyder]

Originally posted by: ArchAngel777
I'll be dissapointed if it is a 256 bit interface. Not that I think the current GTX is memory bandwidth limited, But I do think it would be if the shader power was increased further as these specs might indicate.

With DDR5 it will have plenty of bandwidth. My card shows just under 150GB/s on my 4870 @ 1150MHz, this claims that they'll use 1250Mhz. There would probably be little gain by going 512 bit with DDR5 but I'm sure the costs to build the card go up.

 

[ArchAngel777]

Originally posted by: SlowSpyder

Originally posted by: ArchAngel777
I'll be dissapointed if it is a 256 bit interface. Not that I think the current GTX is memory bandwidth limited, But I do think it would be if the shader power was increased further as these specs might indicate.

With DDR5 it will have plenty of bandwidth. My card shows just under 150GB/s on my 4870 @ 1150MHz, this claims that they'll use 1250Mhz. There would probably be little gain by going 512 bit with DDR5 but I'm sure the costs to build the card go up.

The current 240SP has around 140GB/Sec stock and you can easily overclock to get it in the 150gb/sec range, increasing GPU power 60% and the memory bandwidth 10% seems lopsided. I think even the XFX Alpha XXX runs the memory at a whopping 2500 effective, which translates to 160gb/sec. Now, again, I don't think that the GTX is currently starved for it. In fact, I am guessing it has more than enough. But I am curious to see how it will play out when the GPU power is increased 60% or so and the bandwidth is increased a mere 10%.

BUT - this is all speculation. We don't even know if these specs are true or accurate. We don't even know for sure if it will use GDDR5, and if it does, we don't know the speed that it will use.

 

[Cookie Monster]

I dont think the GTX2x0 cards are bandwidth limited since i haven't seen any evidence that suggests so. Even then the GDDR5 can provide more than enough bandwidth.

 

[AzN]

I don't think this would be leaps and bounds better than GTX 2 series though considering all they've done is replace the 512bit bus with 256bit gddr5 and raised the SP count. Looks like Nvidia is fixing their inefficient design if the specs are true. I would have like to see 384bit bus with gddr5 though.

300mm is rather a big shrink. Little bigger than 4870.

 

[Cookie Monster]

The only inefficiency I can think of is nVIDIA's poor performance/mm^2 and very slow DP when compared to RV770. However its still an excellent architecture although its a shame we won't be seeing any DX10.1 parts from nVIDIA.

 

[Hauk]

Originally posted by: Cookie Monster
The only inefficiency I can think of is nVIDIA's poor performance/mm^2 and very slow DP when compared to RV770. However its still an excellent architecture although its a shame we won't be seeing any DX10.1 parts from nVIDIA.

Yea it looks like they're chasing ATI now in terms of squeezing out memory bandwidth and performance/mm^2. ATI showed us 256-bit DDR5 works and makes for an overal more cost effective pcb so that's not a concern.

The real question is, how will it perform? Anyone want to guess at it compared to GTX 280/285??

 

[Cookie Monster]

~60% is the estimate at this point in time if the spec if true of course.

 

[Creig]

1.8 billion transistors? I wonder what initial yields on that are going to look like?

 

[Hauk]

Originally posted by: Cookie Monster
~60% is the estimate at this point in time if the spec if true of course.

This makes stepping up in the next week a little less exciting.

Guess that's how the cookie crumbles..

 

[OCGuy]

Originally posted by: Cookie Monster
The only inefficiency I can think of is nVIDIA's poor performance/mm^2 and very slow DP when compared to RV770. However its still an excellent architecture although its a shame we won't be seeing any DX10.1 parts from nVIDIA.

DX 10.1 who? You know the Windows 7 beta is out there, right?

 

[nRollo]

Originally posted by: Creig
1.8 billion transistors? I wonder what initial yields on that are going to look like?

This aspect of the equation won't matter to me if two things are true:

1. Yields must be sufficient to send one to me.

2. Yields must be sufficient to sell one to everyone who wants to buy one for $650 or less.



Personally, I've given up speculating on the yields and cost to manufacture. To me it looks like supply and demand set the price, and that's the only thing that affects end users.

A whole lot of people speculated "low yields" on the GTX260 and GTX280 would make them "impossible" to sell for much lower than launch price. Then the market set the price like always, NVIDIA remained profitable, and all was right with the world.

 

[ArchAngel777]

Originally posted by: nRollo

Originally posted by: Creig
1.8 billion transistors? I wonder what initial yields on that are going to look like?

This aspect of the equation won't matter to me if two things are true:

1. Yields must be sufficient to send one to me.

2. Yields must be sufficient to sell one to everyone who wants to buy one for $650 or less.



Personally, I've given up speculating on the yields and cost to manufacture. To me it looks like supply and demand set the price, and that's the only thing that affects end users.

A whole lot of people speculated "low yields" on the GTX260 and GTX280 would make them "impossible" to sell for much lower than launch price. Then the market set the price like always, NVIDIA remained profitable, and all was right with the world.

Hehe, gotta agree with you here. But there is certainly nothing wrong with speculating since it might be the largest chip (transister wise) in history!!!

 

[habbakuk87]

http://www.intel.com/technolog...e-silicon/2billion.htm
Intel is first I think.

 

[Creig]

Originally posted by: nRollo

Originally posted by: Creig
1.8 billion transistors? I wonder what initial yields on that are going to look like?

This aspect of the equation won't matter to me if two things are true:

1. Yields must be sufficient to send one to me.

2. Yields must be sufficient to sell one to everyone who wants to buy one for $650 or less.



Personally, I've given up speculating on the yields and cost to manufacture. To me it looks like supply and demand set the price, and that's the only thing that affects end users.

A whole lot of people speculated "low yields" on the GTX260 and GTX280 would make them "impossible" to sell for much lower than launch price. Then the market set the price like always, NVIDIA remained profitable, and all was right with the world.

Supply and demand does have a lot to do with the end price of a product. But the cost to manufacture also counts heavily as well. If it costs $200 to design, produce and market an item, you can be pretty sure it's not going to retail for $200 or less.

 

[MarcVenice]

With a much smaller die size and nvidia's experience to produce gpu's containing 1.4billion transistors leads me to believe yields won't be as much of a problem as it was with gt200.

And, even if 2, hell 4 of the shaderblocks die in the process, still leaves nvidia's with gpu's that have 320 or 256 stream processors. Those should do for the performance segment.

 

[exar333]

Originally posted by: Creig

Originally posted by: nRollo

Originally posted by: Creig
1.8 billion transistors? I wonder what initial yields on that are going to look like?

This aspect of the equation won't matter to me if two things are true:

1. Yields must be sufficient to send one to me.

2. Yields must be sufficient to sell one to everyone who wants to buy one for $650 or less.



Personally, I've given up speculating on the yields and cost to manufacture. To me it looks like supply and demand set the price, and that's the only thing that affects end users.

A whole lot of people speculated "low yields" on the GTX260 and GTX280 would make them "impossible" to sell for much lower than launch price. Then the market set the price like always, NVIDIA remained profitable, and all was right with the world.

Supply and demand does have a lot to do with the end price of a product. But the cost to manufacture also counts heavily as well. If it costs $200 to design, produce and market an item, you can be pretty sure it's not going to retail for $200 or less.

LOL, tell that to the DRAM manuafacturors.

 

[Sunrise089]

Originally posted by: Hauk

Originally posted by: Cookie Monster
~60% is the estimate at this point in time if the spec if true of course.

This makes stepping up in the next week a little less exciting.

Guess that's how the cookie crumbles..

Considering the OP, nice pun

 

[Cookie Monster]

Well CJ (a well known "leaker" for the lack of a better word) has pointed out that according to some of his sources that GT21x cards support DX10.1.

 

[nRollo]

Originally posted by: Creig

Originally posted by: nRollo

Originally posted by: Creig
1.8 billion transistors? I wonder what initial yields on that are going to look like?

This aspect of the equation won't matter to me if two things are true:

1. Yields must be sufficient to send one to me.

2. Yields must be sufficient to sell one to everyone who wants to buy one for $650 or less.



Personally, I've given up speculating on the yields and cost to manufacture. To me it looks like supply and demand set the price, and that's the only thing that affects end users.

A whole lot of people speculated "low yields" on the GTX260 and GTX280 would make them "impossible" to sell for much lower than launch price. Then the market set the price like always, NVIDIA remained profitable, and all was right with the world.

Supply and demand does have a lot to do with the end price of a product. But the cost to manufacture also counts heavily as well. If it costs $200 to design, produce and market an item, you can be pretty sure it's not going to retail for $200 or less.

The problem is we as end users don't get to know the cost per unit of video cards, we just make speculations based on the die size, number of transistors, layers on the PCB, type and amount of RAM, etc.. We have no clue what went into R&D on the chips, what is spent on marketing and other overhead, if discounts are given for contracts and estimated volume, and probably other factors I'm not even considering.

As an old time forum fixture/ evil NVIDIA minion I talk to people in the industry, and not just at NVIDIA sometimes. I don't know as much as the guys that really work in the industry, but I can say things aren't always as they seem from the outside in the computer hardware industry.

I've long had a theory that ATi parts are clocked closer to their limits, which would lower yields, and the 4800 series more than doubles the SPs of it's predecessor, which would lower yields as well.

In any case, good news for all if true and competition for the impressive sounding "Little Dragon".

 

[Rusin]

Rumour war . If current rumours about RV870 and GT212 would hold truth then RV870 would have about the same performance as GT212 has. Difference would be that RV870 is 140mm^2 when GT212 would be about 300mm^2..
AMD would dominate with bigger margin than they do now.