NVIDIA Kepler speculation thread

[A1xLLcqAgt0qc2RyMz0y]

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

Source: https://registration.gputechconf.com/?form=schedule
Change Drop Down date to Wednesday 5/16

 

[Dave Baumann]

As the retail sale in the market of $550 graphics cards would mean a whole lot to the overall bottomline anyway... Crap article in general.
And not really related to your original qoute, except for the authors own perceptions.

And simply reading the press release points out:

GPU revenue was up in a seasonally down quarter, due to higher improved desktop GPU ASP in the channel, offset by seasonally lower game console royalty revenue.

 

[KimB]

You will never seen the 680 in a laptop, the laptop chips high range will be based on lower sku. And in this range, the laptop chips used will be based on 7870 from AMD and 660 from Nvidia. I need say, the turbo mode from Nvidia could on the paper be really interessant on low end chips for laptop, as it ensure a max tdp and the max possible performance, it will be sold like that anyway.. but the problem is how much difference it can do at low clock speed.. a 10% increase on clock when possible will not be so much profitable. It will make a good feeling on paper, but not in reality. Cause for laptop the average TDP in use and the max are really very close...

Yes, there most definitely will be laptops with GTX 680-derived GPU's. They won't be at all common, but they will exist (not yet, though, obviously). See, for example:
http://www.anandtech.com/show/4897/alienwares-m18x-part-1-nvidias-geforce-gtx-580m-in-sli

However, that's not what I was talking about specifically, but rather the overall performance/power consumption ratio of the architecture, which is pretty tremendous compared to AMD's offerings. If nVidia can retain this for the entire lineup, and can ship those mobile chips in time for laptop vendors this fall, then they may well win quite a few OEM deals. Of course, these sorts of issues are not all that computer manufacturers care about, so it will be interesting to see what happens.

Edit: speaking of laptop parts, just saw this on nVidia's website:
http://www.geforce.com/hardware/notebook-gpus/geforce-gtx-675m

I didn't know the parts had already been announced. It definitely sounds like nVidia is committed to having them available by the fall. Will be interesting to see how well these parts do. And it also makes it sound like nVidia's full 6xx desktop lineup cannot be far away.

 

[lanek]

Yes, there most definitely will be laptops with GTX 680-derived GPU's. They won't be at all common, but they will exist.

However, that's not what I was talking about specifically, but rather the overall performance/power consumption ratio of the architecture, which is pretty tremendous compared to AMD's offerings. If nVidia can retain this for the entire lineup, and can ship those mobile chips in time for laptop vendors this fall, then they may well win quite a few OEM deals. Of course, these sorts of issues are not all that computer manufacturers care about, so it will be interesting to see what happens.

You have ever seen AMD or Nvidia use a top range SKU in laptop range ?

No .... 195W and 65W make a lot difference.. Im sorry but you will see a laptop mobile called 680, but this will not be the 680 desktop .

 

[KimB]

You have ever seen AMD or Nvidia use a top range SKU in laptop range ?

No .... 195W and 65W make a lot difference.. Im sorry but you will see a laptop mobile called 680, but this will not be the 680 desktop .

The GTX 680 also isn't a top-range SKU. It's more equivalent to the previous generation's GTX 560 Ti. Big Kepler is yet to come.

But regardless, we'll see. My initial guess would be that we'll have a 680M sometime next year, or maybe that it will be skipped in lieu of the refresh part's equivalent.

 

[Psycho]

Yes, there most definitely will be laptops with GTX 680-derived GPU's. They won't be at all common, but they will exist (not yet, though, obviously). See, for example:
http://www.anandtech.com/show/4897/alienwares-m18x-part-1-nvidias-geforce-gtx-580m-in-sli

580m is a quite underclocked GF104 part.

performance/power consumption ratio of the architecture, which is pretty tremendous compared to AMD's offerings.

Hard to say from the current offerings.. http://www.techpowerup.com/reviews/ASUS/GeForce_GTX_680_Direct_Cu_II/29.html

Edit: speaking of laptop parts, just saw this on nVidia's website:http://www.geforce.com/hardware/notebook-gpus/geforce-gtx-675m

gtx675m is essentially a renamed gtx580m... http://www.notebookcheck.net/NVIDIA-GeForce-GTX-675M.70785.0.html

 

[Mintmaster]

However, that's not what I was talking about specifically, but rather the overall performance/power consumption ratio of the architecture, which is pretty tremendous compared to AMD's offerings.

You obviously haven't been paying attention to the GPU market. AMD's 7800 series (Pitcairn) is perf/W beast, and also far closer to the TDP usable in a notebook than the GTX 680. GF100/110 had no bearing on NVidia's mobile capability, and the same applies to Tahiti.

Further, I think NVidia will have a tough time matching Cape Verde $ for $ and W for W, which I think will be the fastest GPU that we'll see in a ~4lb notebook this generation.

EDIT:

gtx675m is essentially a renamed gtx580m... http://www.notebookcheck.net/NVIDIA-GeForce-GTX-675M.70785.0.html

And that pretty much confirms that the 680m isn't going to be based on GK104. They wouldn't choose 675m as the name if it wasn't fairly close to the 680m in performance, so we're looking at GF114 (desktop) performance at best for NVidia's top mobile part.

 

[DavidGraham]

Charlie's credibility seems inversely related to how well he talks about Nvidia. When he was pumping GK104 for that little spell there he suddenly became the world's most reliable source on forums it seemed. Over at Hardocp people would be saying stuff like "Charlie proved right again" about minor things...even though as I pointed out, he was still getting almost everything wrong (like, release dates)

Not even when he talks well about NVIDIA , remember the PhysX hardware ?

Completely aggree, when Charlie wrote: " Kepler will be a win", all was praise the Charlie words.. "

He certainly was not the first one to talk about it , we knew that already from Chinese sources , the awe here came from AMD fanatics and Charlie cultists who couldn't believe the sudden turn of disposition .

And really , no one praised him for his idiocy , he proved that he can put logic even in the most illogical technical facts , and still have some people consider him semi-accurate .

 

[trinibwoy]

580m is a quite underclocked GF104 part

GK104 is smaller and more power efficient than GF104. That comparison actually supports Chalnoth's case.

 

[AlphaWolf]

675m is a GF114. Just dial the clock down to 620Mhz (from 825).

 

[AnarchX]

560M: 192SPs@1550MHz
580M: 384SPs@1240MHz

660M: 384SPs@835MHz
680M: 768SPs GK106 would be more suitable, than >1000SPs GK104

Maybe GK104 will be 690M or flag-ship of a late 2012 700M series.

 

[Kaotik]

So, we now know that GK100/110 is 7 billion transistors - roughly double that of GK104.

I'm assuming it will keep the basic ALU-architecture intact - meaning separate FP32-only and FP64-only ALUs

We also know that GK104 SMX with 192 32bit ALUs and 8 64bit ALUs takes 16mm^2
Assuming rest of the SMX stays the same, how much you reckon it would take with 2:1 aka 192:96 FP32:FP64 ALU configuration?
(I know this is really, really simplifying things, it doesn't go that simple, but still might give some idea on how the chip will turn out to be)

 

[3dilettante]

So, we now know that GK100/110 is 7 billion transistors - roughly double that of GK104.

I'm assuming it will keep the basic ALU-architecture intact - meaning separate FP32-only and FP64-only ALUs

I don't think that can be assumed.
The separate and limited FP64 units of GK104 were a result of Nvidia wanting to strip DP capability down while being minimally compatible the CUDA featureset.

The big chip is going to have DP throughput as a priority, and the idea of devoting area to N DP units and 2N separate SP units sounds wasteful.
A multi-precision unit is fatter than a standard one, but the costs of having 3 separate units for a 2:1 ratio versus one unit that can do both are higher.
This inefficiency isn't evident for GK104 because the extra DP elements are so stripped down.

 

[trinibwoy]

Yup there's no way big Kepler will have ~2TF of dedicated DP hardware sitting idle while gaming.

 

[tunafish]

Assuming rest of the SMX stays the same, how much you reckon it would take with 2:1 aka 192:96 FP32:FP64 ALU configuration?

That wouldn't actually be very good for compute. One of the biggest shortcomings (for GPGPU) of the GK104 chip is really low cache/execution unit ratio. They could of course up the caches for the big part, but frankly, it's probably easier to have less execution resources per GPC and have more of them. 48 DP units in a SMX with the present caches would probably be the most that can be served properly.

 

[UniversalTruth]

Posted, no? GTX 690.

Link

 

[Kaotik]

Posted, no? GTX 690.

Link

I dunno, it was linked to be somehow related to the Alienware picture, which it clearly isn't.

 

[lanek]

Not even when he talks well about NVIDIA , remember the PhysX hardware ?


He certainly was not the first one to talk about it , we knew that already from Chinese sources , the awe here came from AMD fanatics and Charlie cultists who couldn't believe the sudden turn of disposition .

And really , no one praised him for his idiocy , he proved that he can put logic even in the most illogical technical facts , and still have some people consider him semi-accurate .

On that , i completely agree with you... But you can only see the reaction on the different forums was funny...

The GTX 680 also isn't a top-range SKU. It's more equivalent to the previous generation's GTX 560 Ti. Big Kepler is yet to come.

But regardless, we'll see. My initial guess would be that we'll have a 680M sometime next year, or maybe that it will be skipped in lieu of the refresh part's equivalent.

This is the only sku they have availble yet.. And i dont buy the Nvidia words as they have a better card, but have wanted be nice with AMD and dont release it. Big kepler have yet to come... but end of 2012 is not timed with the actual series ... Its a damn marketing tricks... Its like AMD tell you: the HD8000 will come and is bigger and faster....

Lets be honest, Nvidia have a big problem with the 600 series. I dont know if it is only production on 28nm.. but we are soon in may, and they have only the 680 with limited availablity, and maybe next month a 690... There's a big problem somewhere, the availibity of the 680 was really limited ( it start to be better now ), and this is the only new carda available from them.. AMD have 6 cards out from entry to high end..
When Nvidia have launch Fermi, they had at least in may the 470-480 out... I Like Nvidia, but this time i dont buy their marketing tricks...

 

[Cookie Monster]

Posted, no? GTX 690.

Link

Looks like its been photochopped by extending the GTX680s backside. I mean the cooling behind the thing doesn't even make sense

 

[psolord]

Looks like its been photochopped by extending the GTX680s backside. I mean the cooling behind the thing doesn't even make sense

I cannot argue about the photoshopping really, but why isn't the cooler making sense?

It's basically the same principle that was used in the GTX 590 and Radeon 6990.