NV40 wishlist?

[Jamin]

Drop the GeForce name and call it "Dawn's Nipples" :D I agree with the prior comments about the inclusion of Dawn pinups etc.

I don't really know enough about the underlying technology to give any serious tech-spec wishlist. But I would like it to at least look cool. With the popularity of case modding, I think more manufacturers should give more thought to the way the card looks. nV should make the reference card look really cool, so that the board makers who do go the extra step will have ultra-l33t looking pieces of art, and those who don't will still at least have something nifty. :)

---Jamin

[ASCI Blue]

An release that's one time. :angel:

[ellroy80]

Mufu,

Any word on improvements in FSAA quality/speed for NV35/40? Any ideas if they are changing the underlying tech.? (I.e. going to TBR or eDRAM)

Can someone also explain to me what the PPP does (in layman's terms)? I understand a bit of the underlying tech, but nothing like Chalnoth, MuFu or Uttar. Thanks.

[Chalnoth]

The idea of a programmable primitive processor is just to give developers a power tool for making use of higher-order surfaces.

Higher-order surfaces are a technology that allows a developer to define a patch with some sort of curvature, and let the hardware generate the triangles that make up the final patch.

There are a number of benefits to using higher-order surfaces:

1. Geometry compression. As geometry counts increase, video memory bandwidth and AGP bandwidth can be a significant limitation to geometry throughput. HOS naturally compress geometry.

2. Scalability. By being able to easily vary the number of triangles rendered, simply by changing a switch when rendering, there is much greater scalability over a wider range of T&L hardware. This can allow for optimum usage of the geometry power of all levels of a generation that supports such technology.

3. Usability. Past architectures have had support for higher-order surfaces, but the support has not been good. The GeForce3/4's patches apparently required far too much CPU power to be usable. The Radeon 8500's Truform technology isn't flexible enough. A robust programmable primitive processor promises to solve these problems, allowing for a flexible HOS technique that also has high performance.

How would a programmable primitive processor be used? There are three possibilities that I know of currently:

1. Patch-based rendering: Similar to the GeForce3/4's RT-patch technology. The essence of the tech is that a patch is defined with vertices and control points. The hardware then interpolates between those control points, in some fashion, to create the final curve. A programmable primitive processor would be able to use an arbitrary interpolation technique (given enough hardware resources).

2. Hardware Displacement Mapping: Could be used quite well for adaptive displacement mapping, similar to that seen in the Matrox Parhelia.

3. Subdivision surfaces: Similar to ATI's Truform, but with designer control over how the new triangles are created, eliminating the current problems with Truform entirely.

Side note:
A programmable primitive processor may actually do nothing but create new triangles. All vertex perturbation on the newly-created triangles could be done in the vertex processor.

[ellroy80]

Thanks Chalnoth, that makes sense.

[Cotita]

Instead of a dual chip or dual card, why not a dual core GPU?

A new AA algorithm that only softens polygon lines, not textures, ala matrox.

Liquid nitrogen cooling :D :D :D

[Chalnoth]

Originally posted by Cotita
Instead of a dual chip or dual card, why not a dual core GPU?
There's not much difference between this and what GPU's do today. They're already highly-parallel.

A new AA algorithm that only softens polygon lines, not textures, ala matrox.
Um, so does the GeForce3, GeForce4, GeForce FX, and Radeon 9500/9700.

As for only storing extra per-pixel data for edges, I'm not sure that can ever work 100% flawlessly.

[Cotita]

Originally posted by Chalnoth
There's not much difference between this and what GPU's do today. They're already highly-parallel.

So are normal CPUs and as far as I know both AMD and Intel are looking into multiple core cpus as a possibility.

Originally posted by Chalnoth
Um, so does the GeForce3, GeForce4, GeForce FX, and Radeon 9500/9700.

As for only storing extra per-pixel data for edges, I'm not sure that can ever work 100% flawlessly.


The GeForce3, GeForce4, GeForce FX, and Radeon 9500/9700 antialiase the whole screen not the edges only.

I'll be happy if it works 90% flaslessly, neither SuperSampling nor multisamplig work 100% flawlessly either.

[Chalnoth]

Originally posted by Cotita
So are normal CPUs and as far as I know both AMD and Intel are looking into multiple core cpus as a possibility.
No, the x86 instruction set does not lend itself well to parallelism. While some paralellism is possible, a very large number of calculations are calculated in serial.

How else do you think the Pentium4 could stand a change against the Athlon? The Athlon has 3 floating point units (though I do forget how general they are, the argument still applies), while the Pentium4 has only one.

By switching to a multiple core system, CPU manufacturers can leverage existing SMP support for better parallelism. This is why, for example, Intel released their "HyperThreading" technology. It is also meant to use existing SMP support for better parallelism (not with multiple CPU's, but just to make better use of what parallelism exists within the CPU).

The GeForce3, GeForce4, GeForce FX, and Radeon 9500/9700 antialiase the whole screen not the edges only.
The algorithm is applied to the entire screen, but the only effect in the output is at polygon edges (not including nVidia's filtered modes, Quincunx and 4x9).

[Cotita]

So why are textures blurred?

[Chalnoth]

Because you're thinking of Quincunx or 4x9. Textures are blurred because color data is shared between pixels (for better edge AA).

[Cotita]

In Quincunx and 4x9 texture bluring is much more noticeable, but all antialiasing modes blur textures, actually they blur the whole screen.

[Uttar]

You *could* consider that MSAA blurs the edges between polygons, and that does mean it can cause some important blurring with very small polygons.

But actually, it isn't really blurring them - it's giving it a color which is nearer to what it should be. Of course, it is not as accurate as with SuperSampling.
I hardly see how that sort of blurring is not desirable...

The reason Matrox's FAA looks good in those situations is that they use 16x *SuperSampling* for edges.
It's not because they only antialiase edges.

So, you're right. But your eye is gotta be darn efficient to be able to notice that much! But yes, Matrox's solution theorically produces slightly less blurring.


Uttar

[Chalnoth]

Originally posted by Cotita
In Quincunx and 4x9 texture bluring is much more noticeable, but all antialiasing modes blur textures, actually they blur the whole screen.
No, they don't. Try taking some screenshots. The only difference on a GeForce 3 (or higher) using 2x or 4x FSAA, or on an R3x0 is on the edges.

These multisampling modes produce absolutely no blurring, because blurring indicates a reduction in color clarity. In terms of MSAA, this means that color data must be shared between pixels. This happens with Quincunx and 4x9, but not any other multisampling modes.

[-=DVS=-]

Moust images will be a bit blurred if you have NO Anistropic Filtering enabled

Offcourse Qunixun is total crap

[Chalnoth]

Originally posted by -=DVS=-
Moust images will be a bit blurred if you have NO Anistropic Filtering enabled
No more blurred than without any FSAA.

[Paul]

Oh god someone just take some screenshots already, else this is going to keep going back and forth.

[Chalnoth]

Fine. This will only be up for a short time (I'll be leaving in an hour or so), but here are the shots. There is clearly zero difference between the shots on any textures that are not in motion.

No FSAA (http://66.127.115.96/Q3NoFSAA.jpg)
2x FSAA (http://66.127.115.96/Q32x.jpg)
4x FSAA (http://66.127.115.96/Q34x.jpg)

[Slappi]

Originally posted by Chalnoth
Fine. This will only be up for a short time (I'll be leaving in an hour or so), but here are the shots. There is clearly zero difference between the shots on any textures that are not in motion.

No FSAA (http://66.127.115.96/Q3NoFSAA.jpg)
2x FSAA (http://66.127.115.96/Q32x.jpg)
4x FSAA (http://66.127.115.96/Q34x.jpg)


Yep you are right and he is wrong.

[Paul]

Can anyone actually see much difference between the 2x and the 4x shots? Whilst there is a substantial improvement in the 2x over no-AA, there doesn't appear to be anything significantly (in that shot, anyway) better to justify the speed cost of 4x.

[cokaznrebel]

hmm...id like to see a video card with a socket(like a mobo socket....) where u can choose what GPU u wanna put in there, they be some cooo shiat, and a couple of RAM slots or somethin...lol...a mini motherboard on the AGP slot..

[abb]

NV40 wishlist? That is simple- Performance and visual quality which Nvidia right now does not have. Hmmm...maybe Nvidia can plant some spys over at ATI, and then they can see what they are doing.:D
Abb

[borntosoul]

will it be using gddr3 ?????
here is a quote , "GDDR3 will consume half the power of GDDR2 and operate up to 50% faster, according to Terry Lee, executive director of advanced technology and strategic marketing at Micron, Boise, Idaho. "

"Lee said Micron's GDDR3 chips are made with 0.11-micron processing, allowing speeds that could reach 700 MHz for a 1.5Gbit/s data rate. The device will sample next quarter, he said."

look good here is the whole link .
http://www.siliconstrategies.com/story/OEG20030321S0036