NVIDIA Kepler speculation thread

[Man from Atlantis]

It's clear that its bullshit, however it does in fact state 3.4b for GK104 and for GK110 6.4b transistors.

IF it is then it's obvious that GK110 is HPC oriented and GK10X for casuals.. at least At Fermi GF114 has 65% transistors of GF110 and this ratio decrease to 53% for Kepler..

6.4b>3b 2.14x
3b>1.4b 2.14x.. question is will GK10X parts have lower performance at HPC tasks compared to GF110? GK104 may be %30 faster than GF110 in games but at HPC jobs it's on par or slightly slower..

personally i expect GK104 is as fast as GTX560Ti SLI(GF114)..

 

[ninelven]

I wonder, would you disable 25% of a 550mm2 chip to hit performance 10% above your own 290mm2 chip?

10% more improvement? It has +30% bandwidth, ~50% more FLOPS, ~50% more texture fill, and ~50% more pixel fill, but it is only going to perform 10% better? I do not understand.

 

[Arty]

10% more improvement? It has +30% bandwidth, ~50% more FLOPS, ~50% more texture fill, and ~50% more pixel fill, but it is only going to perform 10% better? I do not understand.

According to Lenzfire, the performance estimates for 660 is ~580 and +10% for 660Ti. (7950=580)

 

[ninelven]

Well then they have clearly made a mistake somewhere for a chip that is at least 30% faster by every measure to only achieve a 10% performance advantage.

 

[AlphaWolf]

maybe the physx units in the 660ti are crippled

 

[DarthShader]

The GK104 looks like a GF110 die shrink with less memory channels, less ROPs and bumped clocks to make up for the loss. No idea how they would expect the GF106 card with those spec to match a GTX560. Biggest WTF to me is only 1.75GB for the GTX690. That is going to be a bottleneck in high resolutions. Not to even mention the TDP of the whole thing.

 

[Ailuros]

Are those correct?

One at least sounds close enough, but then again considering the ton of mistakes in those tables getting close in just one or few details isn't exactly something that you could call credible. Other than that those must be uber-SPs that 512 with all the bells and whistles account for over 3b and 1024 for over 6b.

According to Lenzfire, the performance estimates for 660 is ~580 and +10% for 660Ti. (7950=580)

Funny doesn't it seem curious that the supposed GK110 has suddenly 3 salvage parts besides the top dog and GK104 only 1?

The GK104 looks like a GF110 die shrink with less memory channels, less ROPs and bumped clocks to make up for the loss. No idea how they would expect the GF106 card with those spec to match a GTX560. Biggest WTF to me is only 1.75GB for the GTX690. That is going to be a bottleneck in high resolutions. Not to even mention the TDP of the whole thing.

2*1.75 = 3.5GB but that doesn't mean that it isn't complete bullshit.

 

[iMacmatician]

I dont know why the GK104 is roughly 580 level when it's projected clocks by Lenzfire for core and shader are both exactly 16.5% higher.

And the 650 is supposedly ~560 (non-Ti) level, yet the former has about 85% of the GFLOPS of the latter. The combination of those two seem quite odd.

 

[Ninjaprime]

But that's quite unlikely, IMO

edit:
To clarify - this is assuming CUDA performance translates more or less directly to gaming performance

Generally you'd see peaks in CUDA that you wouldn't in games and enhancements from their compute focus that won't carry over, so likely gaming percentage would be lower. For example, they can claim twice the performance or more with fermi vs GTX 280 in many CUDA apps, but you probably can't find anywhere in gaming where that is the case. I'd doubt it would translate into more than 20% in games.

Which puts it in the same area as tahiti performance wise. Launching 4 months late at the same performance, price better be damn good...

 

[Ailuros]

Wasn't Lenzfire the site that claimed 164mm2 for Cape Verde and 900MHz for the 7950?

Just for the record you might want to wonder about the buswidth of the GK106 and where the hell GK107 vanished to.

 

[iMacmatician]

Just for the record you might want to wonder about the buswidth of the GK106 and where the hell GK107 vanished to.

4Gamer claimed GK107 may be a disabled GK106 with PCIe 2.0 only. Now, there's already a disabled GK106 (GTX 640) in the LenzFire spec list, but I think there are a few ways to reconcile the disparity:

1. There could be another further disabled GK106 not listed that is a GK107. It could be as minor as a GTX 640 without PCIe 3.
2. The GTX 640 could actually be a GK107, but sources could have assumed it was a GK106 because they are the same chip.

There's a couple other strange things in the table, the presence of nothing but "GTX" parts and the identical core/shader clocks for all three GK110 variants (two is no surprise (GTX 465/470), but three?).

 

[trinibwoy]

Launching 4 months late at the same performance, price better be damn good...

Lol....

Btw, the launch date of the 7970 doesn't dictate how "late" Kepler is. I'm quite sure project managers at the two companies don't share notes.

 

[Psycho]

Ehm.. 6.4B
GF110 is 3B in 530mm2 - is the silly-season-number-maker assuming perfect scaling or just a 600mm2+ chip on a brand new process? (using Tahiti numbers we end around the 550, but usually amd's chips are more dense or just measured differently).
Not to mention the power usage of that GTX690 card - larger chip, smaller process, higher clock => 500W+

And yes, the performance difference between the GK110 parts seems exagarated.
670->680 should be around 15% assuming perfect scaling (which won't be the case) and
the same with the 660ti->670, not those magic 25% steps.
Or IOW, if we believe the 660ti (just to take some reference) the 680 will be more like 20% (marketing %) faster than 7970.

 

[Arty]

Btw, the launch date of the 7970 doesn't dictate how "late" Kepler is. I'm quite sure project managers at the two companies don't share notes.

Where it actually matters, for consumers, it is.

Wasn't Lenzfire the site that claimed 164mm2 for Cape Verde and 900MHz for the 7950?

Yes, they also claimed Jan 9th for 7950 launch.

One at least sounds close enough, but then again considering the ton of mistakes in those tables getting close in just one or few details isn't exactly something that you could call credible. Other than that those must be uber-SPs that 512 with all the bells and whistles account for over 3b and 1024 for over 6b.

Thanks. Are you expecting the SPs to get more area-efficient, compared to Fermi? (Is this an extension of the improved DP perf/W expectation for Kepler?)

 

[trinibwoy]

Where it actually matters, for consumers, it is.

This must be the only industry where all market participants are expected to magically release new products on the same day/week/month. This "late" concept is complete drivel. By that logic if both companies are late then neither are late.

You're late when you miss your internal targets, which nVidia obviously has. However that's got nothing to do with Tahiti's launch date. In other words, by April they will be more than 4 months late.

 

[Arty]

This must be the only industry where all market participants are expected to magically release new products on the same day/week/month. This "late" concept is complete drivel. By that logic if both companies are late then neither are late.

You're late when you miss your internal targets, which nVidia obviously has. However that's got nothing to do with Tahiti's launch date. In other words, by April they will be more than 4 months late.

On the flip side, the internal target makes no sense what so ever. What if your internal target is extremely optimistic/pessimist? The consumer doesnt care if a PM at AMD/Nvidia made an error or had completely lax expectations and let the competition enjoy the market for a significant period.

You can only judge by market opportunity. I doubt either Nvidia/AMD will be as forthcoming as to revealing their internal targets accurately to the public. Most of it is probably dependent on the availability and performance of the new node. From there, its a race to launch.

By your definition, R600 was never late: http://www.anandtech.com/show/2231

While AMD will tell us that R600 is not late and hasn't been delayed, this is simply because they never actually set a public date from which to be delayed. We all know that AMD would rather have seen their hardware hit the streets at or around the time Vista launched, or better yet, alongside G80. But the fact is that AMD had quite a few problems in getting R600 out the door.

One is drivel, the other is not.

 

[trinibwoy]

On the flip side, the internal target makes no sense what so ever.

So internal goals, strategies, budgets and targets are irrelevant? Yet some imagined horse race to see who can "launch first" is the true measure of success? That implies that "launching first" is the primary driver in strategic planning and supercedes all other considerations, which is of course untrue.

It also implies that the consumer cares more about who launches first than the product itself or its availability when they're ready to buy. Im quite sure the entire 28nm generation of GPUs wont be sold in Q1 2012. The bottom line will speak for itself.

By your definition, R600 was never late: http://www.anandtech.com/show/2231

How do you figure that? By my definition it was very late. Just as Kepler is late. A lack of public announcements doesn't give you a free pass on missing internal deadlines.

One is drivel, the other is not.

That we agree on

 

[mczak]

Ehm.. 6.4B
GF110 is 3B in 530mm2 - is the silly-season-number-maker assuming perfect scaling or just a 600mm2+ chip on a brand new process? (using Tahiti numbers we end around the 550, but usually amd's chips are more dense or just measured differently).
Not to mention the power usage of that GTX690 card - larger chip, smaller process, higher clock => 500W+

Well the article states 550mm². Seems at least plausible, though power would be my biggest concern as well (at the high end). All these cards look like essentially shrunk GF110 (with different amount of units of course), and twice the complexity AND higher clock looks a bit like too much even if 28nm gives you 40% better transistor performance. It might be doable though if nvidia got some turbo-like power management.

And yes, the performance difference between the GK110 parts seems exagarated.
670->680 should be around 15% assuming perfect scaling (which won't be the case) and
the same with the 660ti->670, not those magic 25% steps.

Actually the difference between 670 and 680 as stated is only 21% not 25% (since "20% faster than HD7970" and "45% faster than HD7970" does not translate to 25% between those cards).
Still this is higher than expected, since the GTX680 (at the same clocks) would only have 14% theoretical advantage in shader/tmus/rops compared to GTX670. It would not be impossible though if it turns out to be very bandwidth sensitive (which was definitely not the case for GTX5xx) as with these specs it would have 25% more bandwidth.

In any case I don't believe these numbers yet, and especially for the parts which aren't supposed to launch in April it seems extremely unlikely nvidia has finalized clocks for them already.
And nvidia using all high-speed (5Ghz to 5.8Ghz, higher than anything amd shipped yet) gddr5 when their previous record was just 4Ghz? I believe it when I see it...

 

[Arty]

So internal goals, strategies, budgets and targets are irrelevant? Yet some imagined horse race to see who can "launch first" is the true measure of success? That implies that "launching first" is the primary driver in strategic planning and supercedes all other considerations, which is of course untrue.

Completely irrelevant. Launching first is not a success but giving the competition a free pass for months is a failure. It is as irrelevant as is the die size of the GPUs, the consumers dont care how much profit an IHV makes as long as they get it for a price that is acceptable to them.

It also implies that the consumer cares more about who launches first than the product itself or its availability when they're ready to buy. Im quite sure the entire 28nm generation of GPUs wont be sold in Q1 2012. The bottom line will speak for itself.

Exactly. I am pretty sure quite a few 7970 purchases were made by former 580 owners and given the choice would have preferred to buy Nvidia's next over the 7970. No knock on the 7970, its just people preferring one brand over the other.

Bottom line wont matter as in the case of the GF100, it was late and yet it didnt matter. Doesnt mean it met Nvidia's internally targetted date.

How do you figure that? By my definition it was very late. Just as Kepler is late. A lack of public announcements doesn't give you a free pass on missing internal deadlines.

No. AMD never admitted the internal date for R600. By your definition, a lot of the fumbles were actually success, go figure.

That we agree on

Moving goalposts and such.

 

[CarstenS]

In my books, AMD is the clear winner from HD7900/Kepler launches so far - both financially and mind-share wise. From their business point of view they did good on all fronts.

- They have the advantage that their flagship is compared to year-old-tech on an older mfg. process, which makes a difference compared to what Kepler must do: compete with a solution on the same technology basis
- They have the fastest single-GPU card and can command prices they didn't dare to ask for years and at the same time keep demand roughly in balance with supply, which would have been difficult otherwise.
- They can use their solo-time at the top in order to get a re-spin if necessary for tighter binning and maybe better yields - something Nvidia apparently has chosen to do before the actual launch.
- They can utilize the halo effect on their lower end products

The only downside to this approach I can come up with is that some people might not be amused when competition finally drives prices down, but honestly: That's life and it has always been that way (just look at summer 2008 for example).