NVIDIA Kepler GPU Speculation thread

[AnandThenMan]

Now that AMD has dropped the curtain on their latest effort, time to discuss what Nvidia has cooking for us to replace Fermi. Nvidia is said to have working silicon in house, and the so far expected launch is sometime in Q2, 2012.

What does Nvidia have in store to answer AMD? Let the speculations commence.

 

[Flipped Gazelle]

Nice graphic. Did you make it? Looks like you forgot to promote the "vs" object to the front.

 

[96Firebird]

I'm going to guess it used to say ATi, and someone edited over with AMD's logo. Either way, I like the Christmas colors! :thumbsup:

 

[Spikesoldier]

wow havent seen spy vs spy in a long time

 

[AnandThenMan]

Can't take credit for the original, this image has been around a long time, I just swapped out the ATI for AMD Radeon. (still sad to see the ATI name go).

Image fixed.

 

[thilanliyan]

I predict it will have 1024*pi shaders...a 512sin(45) memory bus...and consume 1kwh.

 

[RussianSensation]

Rumor mill:

"GK104 will be a high-performance (≠ high-end) GPU by NVIDIA that will have many of the features that were reserved for its previous high-end GPUs (such as a 384-bit wide GDDR5 memory interface), but will not be NVIDIA's most powerful GPU in the series. The throne will be kept empty for GK100, which will comply with NVIDIA's "go all in" design ideology for high-end GPUs.

3DCenter.org compiled a few specifications of the GK104 and GK100. They go like this:


GK104

• 640 to 768 CUDA cores
• 80 to 96 TMUs (depending on what the CUDA core count ends up being)
• 384-bit GDDR5 memory interface, 48 ROPs
• Built on the 28 nm TSMC process
• Products based on this will launch in the first quarter of 2012

GK100

• 1024 CUDA cores
• 128 TMUs
• 512-bit GDDR5 memory interface, 64 ROPs"

Source

I think GK104 will compete with HD7970 favourably but GK100 will be at least 25% faster than HD7970.

 

[nismotigerwvu]

Assuming Nvidia is willing to make another go at a 500mm^2 , a 6 billion transistor 1024sp monster is entirely possible. Doesn't seem all too likely, but it certainly would be a screamer if they could get decent yields. I think 768 sp (assuming the layout is the same per GPC) at a ~400mm^2 die is more reasonable.

 

[badb0y]

GK104 will be a high-performance (≠ high-end) GPU by NVIDIA that will have many of the features that were reserved for its previous high-end GPUs (such as a 384-bit wide GDDR5 memory interface), but will not be NVIDIA's most powerful GPU in the series. The throne will be kept empty for GK100, which will comply with NVIDIA's "go all in" design ideology for high-end GPUs.

3DCenter.org compiled a few specifications of the GK104 and GK100. They go like this:
GK104

• 640 to 768 CUDA cores
• 80 to 96 TMUs (depending on what the CUDA core count ends up being)
• 384-bit GDDR5 memory interface, 48 ROPs
• Built on the 28 nm TSMC process
• Products based on this will launch in the first quarter of 2012

GK100

• 1024 CUDA cores
• 128 TMUs
• 512-bit GDDR5 memory interface, 64 ROPs

Eagerly awaiting GK100 unless it is 800$ lol

 

[toyota]

Rumor mill:

GK104 will be a high-performance (≠ high-end) GPU by NVIDIA that will have many of the features that were reserved for its previous high-end GPUs (such as a 384-bit wide GDDR5 memory interface), but will not be NVIDIA's most powerful GPU in the series. The throne will be kept empty for GK100, which will comply with NVIDIA's "go all in" design ideology for high-end GPUs.

3DCenter.org compiled a few specifications of the GK104 and GK100. They go like this:

GK104

• 640 to 768 CUDA cores
• 80 to 96 TMUs (depending on what the CUDA core count ends up being)
• 384-bit GDDR5 memory interface, 48 ROPs
• Built on the 28 nm TSMC process
• Products based on this will launch in the first quarter of 2012

GK100

• 1024 CUDA cores
• 128 TMUs
• 512-bit GDDR5 memory interface, 64 ROPs

Source

I think GK104 will compete with HD7970 favourably but GK100 will be at least 25% faster than HD7970.

really? even a 768sp GK104 would be around 25% faster than a 7970. a 7970 is only 20-25% faster than a gtx580 and a 768sp GK100 would probably be 50% faster than a gtx580.

 

[MrTeal]

Eagerly awaiting GK100 unless it is 800$ lol

If Kepler has the performance most people seem to think, a 1048 CUDA version might be able to command $800, and depending on what yields are at that point the problems with getting enough fully functional dies at 500+mm^2 might demand that kind of price.

It might not be that bad, but it could be. If they do charge that much, I'd at least hope they'd call it the 680 Ultra. It would just feel right.

 

[Cookie Monster]

I thought 768CCs would be the next logical step (something along the lines of double GF114 along with architectural optimizations/tweaks/changes seeing as kepler is focusing more on Perf/W where as maxwell will be the next big thing) but rumours are floating around about a ~1000CC part (GK104) and ~1500CC part(GK100).. and getting rid of shader clock but instead upping the base core clock higher.

With nVIDIA looking to release bottom to top instead, im guessing we might see GK104 in january and then see the big thing later in Q2/12 which seems to coincide with the 6 month delay rumour.

It sure will be interesting to see what they have in store since GCN is a solid competitor.

 

[AnandThenMan]

I wonder if TSMC has a love/hate relationship with Nvidia. The bean counters see the volume and do a happy dance, the engineers see the transistor count and head for the bottle.

I'm more interested to see what Nvidia is going to be doing on the compute side of things myself.

 

[RussianSensation]

really? even a 768sp GK104 would be around 25% faster than a 7970. a 7970 is only 20-25% faster than a gtx580 and a 768sp GK100 would probably be 50% faster than a gtx580.

I am being conservative because I don't want people to point out to me later that you can buy a factory pre-overclocked HD7970 with 1.1ghz clocks. I also have to account for HD7970 gaining about a 10% increase in performance through drivers. :biggrin: Not that NV won't add more performance through driver increases (so it might be a moot point). They just added 25% more performance for SKYRIM (indoors).

I am almost shocked that HD7970 is only 25-30% faster than a GTX580 because NV added 15-18% just by going from GTX480 to 580 without even trying. :hmm:

Still I will choose to remain conservative for Kepler because those specs are just rumors, not facts. For all we know GK100 might have 768 SPs and the 1024 version is 2x GK104s? Still, AMD set the bar really low this time.

 

[tviceman]

really? even a 768sp GK104 would be around 25% faster than a 7970. a 7970 is only 20-25% faster than a gtx580 and a 768sp GK100 would probably be 50% faster than a gtx580.

The gtx 560ti is right at 50% faster than the gtx285, so if Kepler can avoid most of the design problems that plagued first gen fermi's, then that is not a bad assumption.

I do not think Kepler's "gtx560ti" will end up quite that fast though. I think it will more or less be equal to the hd7970 in performance, which would then mean Nvidia could potentially release 3 variant GK100 based GPU's, all of which would have higher performance than the hd7970.

I'd also like to say to the OP that Nvidia has never said exactly when kepler wil be out. The roadmap floating around was never verified as true or accurate in any way, so Kepler's release is strictly TBA. It could be Q2, it could be Q3, it could also be Q1. As to which Kepler chip will come first, again, no verifiable evidence in any shape or form has been produced.

EDIT: I also think Kepler's flagship will cost at least as much as the hd7970. Most likely $599.99 for the highest end, $499 for the second tier, and $349-$399 for the third tier GK100 (if there is one). The top GK104 card will be $299-$349. Again, all guesses.

 

[Hypertag]

I am being conservative because I don't want people to point out to me later that you can buy a factory pre-overclocked HD7970 with 1.1ghz clocks. I also have to account for HD7970 gaining about a 10% increase in performance through drivers. :biggrin: Not that NV won't add more performance through driver increases (so it might be a moot point). They just added 25% more performance for SKYRIM (indoors).

I am almost shocked that HD7970 is only 25-30% faster than a GTX580 because NV added 15-18% just by going from GTX480 to 580 without even trying. :hmm:

Still I will choose to remain conservative for Kepler because those specs are just rumors, not facts. For all we know GK100 might have 768 SPs and the 1024 version is 2x GK104s? Still, AMD set the bar really low this time.

You seem to be the only person besides me thinking this out like this. Everyone else seems convinced that no 28nm Nvidia parts will be out until 2013 or 2015, and that the 7970 will be replaced by the 8xxx series in six months.

 

[Nintendesert]

You seem to be the only person besides me thinking this out like this. Everyone else seems convinced that no 28nm Nvidia parts will be out until 2013 or 2015, and that the 7970 will be replaced by the 8xxx series in six months.

2015? :hmm:

 

[AnandThenMan]

I'd also like to say to the OP that Nvidia has never said when kepler wil be out. The roadmap floating around was never verified as true or accurate in any way, so Kepler's release is strictly TBA. It could be Q2, it could be Q3, it could also be Q1. As to which Kepler chip will come first, again, no verifiable evidence in any shape or form has been produced.

Hence the reason for a speculation thread. I think we can all agree on one thing, Nvidia needs to get Kepler out ASAP, or at least have refreshed Fermi parts coming out shorty.

When is Nvidia's next earnings call? That is probably when we will get some concrete-ish info.

You seem to be the only person besides me thinking this out like this. Everyone else seems convinced that no 28nm Nvidia parts will be out until 2013 or 2015, and that the 7970 will be replaced by the 8xxx series in six months.

I think you might be imaging things.

 

[RussianSensation]

You seem to be the only person besides me thinking this out like this. Everyone else seems convinced that no 28nm Nvidia parts will be out until 2013 or 2015, and that the 7970 will be replaced by the 8xxx series in six months.

Don't worry. I was here when AMD beat Fermi by 6 months. The discussions then involved the going-concern for NV, possible bankruptcy, severe loss of desktop discrete market share, loss of brand value, inability to ever come back (being behind AMD forever by at least 6 months), low yields resulting in horrible gross margins, etc. etc. You just have to remember that NV recovered from FX5800/5900 and AMD recovered from HD2900/3800 fiasco. In a nutshell, right now AMD only has 1 videocard that is hands down better than anything NV has, a $550 one at that, out of reach for 95% of the market. The 28nm race is just getting started in my eyes.

I think you might be imaging things.

His point was that a lot of people on our forum have already written-off Kepler because it's "late", expect AMD to refresh HD7970 shortly without problems (like Q2), while expecting 28nm Kepler to be delayed FAR into 2012, with performance increase. My view is not in-line with that sentiment. I hope you caught on to his sacrasm because he didn't literally mean 2015 and HD8000 series in 6 months.

 

[fourdegrees11]

Not to get too far off the Kepler discussion, but I dont expect a 8000 series. I'm looking at the massive OC headroom left by AMD, thinking about as the process matures gaining even more stable higher clocks, and then looking at how AMD released the 4890 in the past. 1.2-3 ghz stock clocked 7980 in 6 months?

 

[Saico]

His point was that a lot of people on our forum have already written-off Kepler because it's "late", expect AMD to refresh HD7970 shortly without problems (like Q2), while expecting 28nm Kepler to bebe delayed FAR into 2012 and for it to unlikely to bring a huge performance increase. My view is not in-line with that sentiment. I hope you caught on to his sacrasm because he didn't literally mean 2015 and HD8000 series in 6 months.

Well, you're not any better than anyone from the hordes of "7970 60% better than 580" fanboys.

Stripping off "rambus and 1024 cores" from GTX780 we'll get another 10-20% faster than 7970 for 200$ more, 6 months later. Get real.

 

[AnandThenMan]

His point was that a lot of people on our forum have already written-off Kepler because it's "late", expect AMD to refresh HD7970 shortly without problems (like Q2), while expecting 28nm Kepler to be delayed FAR into 2012 and for it to unlikely to bring a huge performance increase.

Whether Kepler is late is a relative thing. As soon as the 7970 ships in volume, Kepler could be considered late. I have already said, I will be quite surprised if Nvidia has no answer for the next 5-6 months, it's not like Nvidia didn't know what AMD was doing, they were and are fully aware I'm sure.

I don't really understand why Kepler would be months out, Nvidia already did the hard stuff with Fermi, seems to me a tweaked and updated design would not be nearly as challenging.

 

[badb0y]

Keplar can't be late because it never had a release date.

I am going to buy the HD 7970 just to mess around with it till nVidia releases Keplar .

 

[skipsneeky2]

Whether Kepler is late is a relative thing. As soon as the 7970 ships in volume, Kepler could be considered late. I have already said, I will be quite surprised if Nvidia has no answer for the next 5-6 months, it's not like Nvidia didn't know what AMD was doing, they were and are fully aware I'm sure.

I don't really understand why Kepler would be months out, Nvidia already did the hard stuff with Fermi, seems to me a tweaked and updated design would not be nearly as challenging.

We don't need another g80/g92 hack job for nvidia to cut it.

Amd built its new gpu from the ground up i would assume nvidia will do the same.

 

[Flipped Gazelle]

You seem to be the only person besides me thinking this out like this. Everyone else seems convinced that no 28nm Nvidia parts will be out until 2013 or 2015, and that the 7970 will be replaced by the 8xxx series in six months.

Nobody here has said anything remotely like this. Most people here will be shocked if Nvidia doesn't have Kepler launched within 3-6 months, and a bumped-up 7970 is a very natural evolution.