[Nvida] [Tweaktown] GK110 maybe = 7 Billion Transistors

[n0x1ous]

http://www.tweaktown.com/news/23678/nvidia_s_7_billion_transistor_kepler_gpu_the_gtx_690/index.html

Nvidia's upcoming GTC conference references a 7 Billion Transistor "GPU", not plural, leading to possiblity that upcoming teaser is for GK110 and not GK104 dual GPU card.

What a monster!!!!! Official Nvidia text about the "Inside Kepler" portion below

S0642 - Inside Kepler
Stephen Jones (NVIDIA) , Lars Nyland (NVIDIA)

In this talk, individuals from the GPU architecture and CUDA software groups will dive into the features of the compute architecture for "Kepler" - NVIDIA's new 7-billion transistor GPU. From the reorganized processing cores with new instructions and processing capabilities, to an improved memory system with faster atomic processing and low-overhead ECC, we will explore how the Kepler GPU achieves world leading performance and efficiency, and how it enables wholly new types of parallel problems to be solved.

Topic Areas: Parallel Programming Languages & Compilers
Level: Beginner

Day: Wednesday, 05/16
Time: 2:00 pm - 3:20 pm
Location: Hall 1

 

[nismotigerwvu]

They would need at least 600 square centimeters to pack all those in at current densities. I don't know if I see 7 bill happening. 6 Billions sounds more in line (but still HUGE).

 

[MisterMac]

I can't compute this.

It can't be GK104, as it doesn't compute.


If they release a gk110 in a professional market, why not release gk110 for us?


NVidia stop trolling damnit -_-

 

[BallaTheFeared]

 

[Ben90]

They would need at least 600 square centimeters to pack all those in at current densities. I don't know if I see 7 bill happening. 6 Billions sounds more in line (but still HUGE).

It would end up being about 580mm^2 at current densities. Honestly, I would imagine it would be smaller since a compute oriented card is going to have more L2$. It is very possible that it ends up being just slightly larger than GF110.

Keep in mind that high margins cards won't need to have as good yields compared to the more mainstream GPUs. Nehalem-EX is a good 680mm^2.

 

[MisterMac]

A question would be here... are they really seperating the lines?

If this is Tesla styled line - will we EVER see a consumer version with poor yields going on?

I mean prices are flip 5x on the professional market.
They can stand poorer yields.

Can you imagine a non compute card with 7 billion transistors?
..my loooooooooordy lord.

 

[toyota]

I can't compute this.

It can't be GK104, as it doesn't compute.


If they release a gk110 in a professional market, why not release gk110 for us?


NVidia stop trolling damnit -_-

because they can make way more money per card in the professional market. it makes more sense to do that if they are not producing a whole lot of them.

 

[taltamir]

It would end up being about 580mm^2 at current densities. Honestly, I would imagine it would be smaller since a compute oriented card is going to have more L2$. It is very possible that it ends up being just slightly larger than GF110.

Keep in mind that high margins cards won't need to have as good yields compared to the more mainstream GPUs. Nehalem-EX is a good 680mm^2.

Also there is die harvesting.

 

[Meaker10]

64bit FPUs perform the same as 32bit FPUs in games yet take up more transistors.

 

[AdamK47]

It's great news if the teaser is indeed for the GK110. It's better to have it sooner than than later.

 

[Tempered81]

Sweet one market analyst calculates 3.5bx2 = 7b on the eve of dual-104 product launch, and now 110 is a 600mm^2 die with 7 billion xtors.

 

[tviceman]

7 billion transistors fits right in with current transistor densities both sides are getting at 28nm, and also with the rumored ~550mm^2 die space.

This chip is going to be a monster, and it's also going to cost as much. If it comes out at all on the consumer side, I'm guessing it will debut around $649 and come packed with 4 gig of vram. I'd imagine there would be several harvested parts, probably around $549 and $449, which would then push GK104 based parts to $399 and lower. So basically, I don't think we will be seeing any price movement for the gtx680 anytime immediately before GK110 (unless AMD sparks a real price war).

 

[Tempered81]

I'm thinking it's around 495mm^2 and 5.5b xtors with 3gb/6gb on 384bit, and that this rumor is based on a misrepresentation from a blogger adding 3.5 + 3.5, and calling GTX690 dual gk104 a "GPU".

Just like when GTX680 was supposed to run on a "200" Watt PSU, and the original article stated that 3 GTX580's showing the original Samaritan demo ran on a 2000 watt PSU 1 year ago. 200 was written instead of 2000 in the typo, then a 200-watt rumor started and it turned into "GTX680 ran on a 200 watt psu"

Now it's the same thing in a blog, a guy writes that Nvidia is about to release a product (we all know it's a dual gk104) that product uses 3.5+3.5 = 7b xtor's worth of logic on the board. (Uhh its two dies though). "Hello public, Nvidia on verge of releasing 7 billion xtor graphics product". Now 110 is no longer a 384bit 2304CC X4500 Kepler die, but a 512bit, 3600CC, X6000, 7billion xtor, 600mm^2 die? Give me a break! lol.

Yes two gk104 = 512bit, 7bil xtor, 600mm2, .... However, 110 = 384bit, ~5.5b xtor, ~495mm2, IMO.

 

[Tempered81]

But if I'm wrong and this thing is 7billion xtors... HOLY JESUS

 

[n0x1ous]

Nvidia themselves reference a 7 Bill xtors GPU (singular) in their GTC right up so that certainly points to GK110

plus they mention the compute aspects of it since its GTC which GK104 Lacks big time so again pointing to BIG K

 

[Tempered81]

Agreed. BUT! Nvidia also referenced a 200 watt PSU on their blog about FXAA in Samaritan. Seven bil xtor would be ~600mm2 itanium size at their density, and sounds absolutely absurd and ridiculous. Would also put the GK110's projected X4500 11 score well over X6000+.

Yes Nvidia will make the largest logic they've ever attempted on the poorest yielding most expensive wafers in their history.... No doubt about it. While this blog has no reputation at stake for this particular comment. It's almost not even worth debating, however the slightest hope that it could be true keeps me interested.

 

[Haserath]

Yet Nvidia is also pushing for bigger wafers.

Coincidence?

 

[MisterMac]

If it wasn't for that compute discussion going on, there'd be no doubt about GK110.

Let's maybe say, gk110 will come in consumer form after summer and professional first.

And gk104x2 is still on the table in the mean while.


Logical?

 

[Jaydip]

I think we will see a glimpse of BigK in the upcoming GTC.But this 7B card is probably 690=2X680

 

[MisterMac]

I think we will see a glimpse of BigK in the upcoming GTC.But this 7B card is probably 690=2X680

Except it's not, since gk104 still has NO compute.



Unless they redesigned the entire chip just to make a dual version with compute.


...right.

 

[Jaydip]

Well they aren't going to redesign because i think they have changed their strategy.I hated when they cut the DP arithmetic to 1/24th compared to 1/8th of 580.This strategy will benefit them immensely but some gamers who like to do some compute tasks will be supremely pissed off.

 

[Keysplayr]

Except it's not, since gk104 still has NO compute.



Unless they redesigned the entire chip just to make a dual version with compute.


...right.

What do you mean NO compute? Use Media Show Expresso lately with a 680?
What you're talking about is much lower double precision operations on GK104 compared to a GTX580 for example. Not all compute operations require double precision at all. So, while 680 is weaker in dp, we still have yet to see how it fares in other single precision compute functions. It's a new arch and software is still being worked on for it.

http://www.techpowerup.com/161914/G...ompute-Power-Than-Tahiti-(HD7970)-Report.html

Big K however, looks to be a juggernaut.

 

[Silverforce11]

Whats the leaked specs for gk110, 25% extra CUDA cores, 512b bus, 4gb vram?

All that extra transistors obviously gone to increasing DP/HPC performance.

 

[Keysplayr]

Whats the leaked specs for gk110, 25% extra CUDA cores, 512b bus, 4gb vram?

All that extra transistors obviously gone to increasing DP/HPC performance.

http://www.legitreviews.com/news/12803/

"This latest leak suggests a slightly later release date than thought - not August, but September. This bad boy will feature twice the RAM of the GTX 680, making it a monstrous 4GB RAM, while other specs are as previously reported, 2304 CUDA cores 512-bit memory bus, 250W TDP and performance around 20-25% better than GTX 680. To be honest though, with the double-size memory bus, one would expect a better performance advantage than this, since such a bus is very expensive to implement in real hardware, hence a larger expected payoff is reasonable. These are early days yet, so the performance could well be more like 50% faster by the time it's launched."

http://www.fudzilla.com/home/item/26459-geforce-kepler-gk110-basic-specs-leaked

"It is expected that GeForce Kepler GK110 may have up to 2304 CUDA cores - up from GK104's 1536 CUDA cores - thanks to its bigger die which allows more SMX units (Streaming Multiprocessors x 2) to be placed on silicon. The new chip will also contain a total of nearly 6 billion transistors all drawing a total TDP between 250 and 300 watts."

http://videocardz.com/31126/geforce-kepler-gk110-specification

"According to 3DCenter, new GK110 chip may have up to 550mm2 of die size. This would be 87% more than actual GK104. Some say this is too much, and NVIDIA will not release gpu that large, because cost production of such gpus is significantly higher. But latest rumors suggest that it may be somewhere in those dimensions. Bigger die allows engineers to put more SMX’s on new gpu. It is said that GK110 may have up to 2304 CUDA cores. NVIDIA is already preparing their GeForce GTX 690, a dual-GK104 card, which will have 3072 CUDA cores."

 

[Silverforce11]

Along with the leaks for gk104, the chinese months ago have said that big K gk110 will be nearly 600mm2 at the limits of production capability.. so I think its fairly accurate it will be huge. 20-25% better performance means its clocked quite low in comparison (TDP issue) and the extra transistors don't contribute much to gaming performance.

Given that its the true successor to the gtx580, 50-60% better performance is pretty realistic. Still, Q4 and prolly low yields.. 28nm is so meh...