next gen kepler to support dx 11.1, also take a year to rollout all cards

[ViN86]

One word Nvidia: NetBurst.

Learn from the past.

[Xion X2]

Originally Posted by Viral

We don't know the performance or actual power draw yet so it's a bit soon to say that they are inefficient. Fermi was inefficient on perf/trans and perf/watt compared to Cypress/NI for gaming yes, but much of the design was designed for GPGPU. Now that both are strong in this area a real comparison can be made. There's also the common occurrence of that last 10-5% performance costing much more in transistors and power use - if the top end Kepler does use up to 300w, but performs 20-40% faster than Tahiti, I don't think that's really THAT unreasonable. If it's just 5-15% again, that's not worth it for me.

It's true that we don't know enough yet; that's why I said we don't have "conclusive" data in reference to the new lineup.

What I was speaking to was past history and, more specifically, the past 3 generations of cards. For the past 3 gens, Nvidia has really lagged behind when it comes to efficiency. Their cards have been fast at the cost of consuming a lot more power than ATI's. When you're trying to increase performance, that envelope will run out unless you figure out a way, like ATI has, to increase efficiency within your architecture. Extra transistors come at a cost.

My feeling is that in the end they will figure it out but it may take them some time to get their next high end out there.. like in the latter part of this year.

[shadow001]

It is true that the high end parts for ATI have been within roughly the same max power limit for the last few generations(about 210 to 225 watts), as i don't remember the HD 6970 using more power than the HD5870, or these new HD7970's using more power then the HD6970's under load, and the transistor budget between the HD5870 and HD7970 doubled(2.15 billion to 4.3 billion), and of course the process changed from 40 nm to 28nm, but the die sizes aren't hugely different either, going from 315mm^ to 389mm^ and back to 365mm^ respectively for all 3 generations.


So they stay within reasonable ranges for power and die size, and the common point with Kepler is that it's TSMC mass producing Kepler chips, including the very highest end version using that same 28nm process, so if we start hearing that it's using a much higher transistor budget than AMD's high end GPU, power concerns might possibly pop up unless it runs at lower clocks, but they can't avoid getting less chips per wafer even if yeilds were 100%, as the GPU die is larger too as a consequence of the higher transistor budget anyhow, so production costs are automatically higher for them than for AMD, and Nvidia needs to order more wafers from TSMC to make up the shortfall.

[Muppet]

It's still all just conjecture at this stage as Nvidia's have not released any info about these next chips. All we have seen is rumors about what it could be like. Personally, I'm waiting until Kepler is released or solid info from a reliable source. That's when I'll make up my mind which camp I go with next round.

[MUYA]

Originally Posted by Muppet

It's still all just conjecture at this stage as Nvidia's have not released any info about these next chips. All we have seen is rumors about what it could be like. Personally, I'm waiting until Kepler is released or solid info from a reliable source. That's when I'll make up my mind which camp I go with next round.

With GK104 out all details could possibly make the foundations for better gauging of the GK 110..where are the GK 104s!!!!

[shadow001]

Originally Posted by Muppet

It's still all just conjecture at this stage as Nvidia's have not released any info about these next chips. All we have seen is rumors about what it could be like. Personally, I'm waiting until Kepler is released or solid info from a reliable source. That's when I'll make up my mind which camp I go with next round.

True, but just curious that they're so silent about it and we're almost 2 months since the HD7970 reviews were released on december 9th last year, and one month in actual retail shelves, and best case scenario( if the rumors are to be believed), the first kepler based cards will be released only in april, and it's not the high end version either....


Anyhow, i dug up the power chart under load while playing metro 2033 from the Anantech review of the HD7970 and i was right..It's not really using much more than the HD6970 or HD5870 cards(40 watt spread between the 3), and it's the same system being used apart from the video cards.

[MUYA]

Originally Posted by ViN86

One word Nvidia: NetBurst.

Learn from the past.

I am sure Nvidia learnt from NV30, Fermi as well...

[Muppet]

Originally Posted by MUYA

With GK104 out all details could possibly make the foundations for better gauging of the GK 110..where are the GK 104s!!!!

That's true, when is the GK104 due out. Last I heard it was sometime this month. I have only heard rumors about April for the release of the GK110.

[shadow001]

Originally Posted by Muppet

That's true, when is the GK104 due out. Last I heard it was sometime this month. I have only heard rumors about April for the release of the GK110.

That was Rollo saying it.....Nuff said there really.

[Muppet]

Originally Posted by shadow001

That was Rollo saying it.....Nuff said there really.

I never take anything he says seriously. I read those over at Extreme Systems. Still only rumors though. That was why I was asking MUYA as Asia seems to get hardware releases before anywhere else in the world.

[jlippo]

Originally Posted by Vardant

It was because of displace mapping/height mapping. For example, if you have a 128x128 texture, tesselation needs to smooth the polygon to 128x128 or more polygons.

If you go even further to 256x256, it allows for smoother transitions and the object appears much more plastic, thanks to tessellation and displace/height mapping.

If this is the case they have problem on how they are tesselating the objects.
They should use view dependent tesselation in which case you can use error metrics to decide how many polygons are needed. (by curvature, pixel/polygon size in screenspace and so on.)

Originally Posted by Vardant

The other reason, for almost over-using tessellation is a problem with very thin and long polygons. The GPU has a really hard time with dealing with polygons, that are for example only 1mm wide, but 5m long. It's a problem, that every engine he worked on had. Dynamic tessellation on singlemesh objects should help with that, but Crysis 2 apparently doesn't support it. And because it's missing, the level of tessellation doesn't change with the distance of camera.

This would be a case where view dependent tesselation would help a lot, but art should be made with tesselation in mind as well.

I really hope that we would see a proper case for tesselation in some future game.
IE.
Mesh prepared so that it would have decently even structure, using vector displacement mapping and texturing which has UVWs relaxed properly, so it will not stretch with the dispalcement.

[Viral]

Originally Posted by jlippo

I really hope that we would see a proper case for tesselation in some future game.
IE.
Mesh prepared so that it would have decently even structure, using vector displacement mapping and texturing which has UVWs relaxed properly, so it will not stretch with the dispalcement.

You'll start to see it once the next gen consoles are out....

Crysis 2's tesselation implementation is a complete joke and no one in their right mind should be defending Crytek over this. They released a hacked, unoptimized solution for whatever reason. Even if they couldn't get the dynamic tesselation implemented and couldn't manage to clip the water, there's no damn reason to use that many triangles on a flat surface! Especially when the detailed bits don't benefit from it at all (still simple detail, just more triangles splitting up what's already there for the most part).

GK104 launch date; earliest speculated is mid march, latest is late april. GK110 no one has any clue if it will launch with GK104 or not, but I'm starting to think it won't considering the details about GK100 being scrapped and GK110 rushed forward.