Toms Hardware: GeForce 5200 is 2x2?

[Joe DeFuria]

Lars also talks about some "ballpark" performance:

http://www.tomshardware.com/graphic/200303061/geforcefx-5600-5200-01.html

However, we can already give you some of our first impressions about the performance of the new cards. In the standard tests without FSAA and anitotropic filtering, the FX 5600 Ultra seems to be just about the same or slower than a GeForce 4 Ti 4200 8x and Radeon 9500 PRO. This might be due to the reduced pixel pipelines (2x2 as opposed to 4x2). With 4xFSAA, it appears to reach nearly double the performance of the 4200, beating a 4800 as well, but it loses out to the Radeon 9500 PRO. It's a similar picture with the anisotropic filtering. In the pixel shader tests from 3D Mark 2001, it beats the 4200/2800, but loses in the vertex shader tests. In both tests, it clearly loses to the Radeon 9500 PRO.

The FX 5200 Ultra is quite a bit slower than the 5600 Ultra in the standard tests. In anti-aliasing, it's just a tad behing the Ti 4200. It's the same with anisotropic filtering - at least in Balanced mode. In Performance mode, it even manages to beat the 4800.

On the surface, it looks like the 5600 Ultra and 9600 Ultra will be farily evenly matched....but the 5200 Ultra may have some problems with the 9600 non Ultra...

I repeat again...rush out and get those 9500 Pros before they're all gone.

 

[overclocked]

Can we draw any conclusions besides that the 9500pro seems to be a great deal..!! This is so strange.. ATI has really turned the pan..

 

[overclocked]

Btw is 2*2 equal to 2*2 + 2z? Getting very confused here.

 

[RussSchultz]

Another review stated it was 4 pipelines and specifically mentioned that it was 4 full pipelines, and not the 1+2z thing of the NV30.

I'll wait until more benchmarks come out before putting much stock in anything.

 

[Joe DeFuria]

Agreed...though wouldn't it be nice if when they said "4 pipelines", we would have reasonable confidence in what they meant?

 

[Mulciber]

maybe the nv31 is actually 4x1+2z
that would be a nice bonus for DoomIII engine games

That tomshardware review is very contradictory. I wouldn't even try to get any information out of it. It looks like a very poor translation from their German site.

He says theres a huge gap between the 5200U and the 5600U, but then he goes on to say they both perform around the Ti4200 level.

 

[RussSchultz]

Agreed...though wouldn't it be nice if when they said "4 pipelines", we would have reasonable confidence in what they meant?

I would certainly prefer that.

 

[Joe DeFuria]

OK...I'll stick out my neck and make a guess on the architectures:

NV31: 4x1 + 2z. (4 pixels + 2 Zixels)

NV34: 2x2 + 2z. (2 pixels + 2 Zixels)

Now...we wait!

 

[BRiT]

I thought the NV31 and NV34 did not have the Zixel improvements the NV30 has.

 

[OpenGL guy]

OK...I'll stick out my neck and make a guess on the architectures:

NV31: 4x1 + 2z. (4 pixels + 2 Zixels)

NV34: 2x2 + 2z. (2 pixels + 2 Zixels)

Now...we wait!

It wouldn't make sense to have less zixels than pixels... you could just have your pixel units output the Z values and get better performance. Plus, if you're doing color + Z, then you're Z values would be at half the speed of color for the NV31 you're describing.

My guess:
NV31 = 4x1 w/ 4z
NV34 = 2x2 w/ 2z

All speculation of course as I'm basing my guesses based on preliminary performance numbers that have been posted.

 

[Joe DeFuria]

OpenGL guy...yes I know. I ust stated it differently.

We both have the same exact guess for the NV34. The way I stated it, a "Pixel" can be either a colored pixel, or a zixel. EDIT: I re-read your guess and I realize it's not the same...oh well.) To be clear:

NV34: 2x2 + 2z. (2 pixels OR 2 + 2 = 4 Zixels)

NV31: 4x1 + 2z (4 pixels, OR 4+2=6 Zixels).

The only difference between my guess and yours, is that I'm guessing the NV31 can do a total of 6 Zixels per clock, rather than 4.

 

[OpenGL guy]

OpenGL guy...yes I know. I ust stated it differently.

We both have the same exact guess for the NV34. The way I stated it, a "Pixel" can be either a colored pixel, or a zixel. To be clear:

NV34: 2x2 + 2z. (2 pixels OR 2 + 2 = 4 Zixels)

NV31: 4x1 + 2z (4 pixels, OR 4+2=6 Zixels).

Ok, now I see what you are saying.

The only difference between my guess and yours, is that I'm guessing the NV31 can do a total of 6 Zixels per clock, rather than 4.

Seems very unlikely that the number of zixels would not be an integer multiple of the number of pixels.

 

[Ilfirin]

My god, don't tell me the zixel term is actually catching on?!

 

[Joe DeFuria]

Seems very unlikely that the number of zixels would not be an integer multiple of the number of pixels.

Yeah, but if I get it right, I'll look like a bon-a-fide prohpet...why esle would I guess such a crazy thing.....whereas if you are right..."eh...that was obvious."

 

[OpenGL guy]

Yeah, but if I get it right, I'll look like a bon-a-fide prohpet...why esle would I guess such a crazy thing.....whereas if you are right..."eh...that was obvious."

If I'm right, maybe nvidia will redefine what "nxm" means again just to make me wrong

 

[Ostsol]

In the pixel shader tests from 3D Mark 2001, it beats the 4200/2800, but loses in the vertex shader tests.

What do you guys think the deal is with the vertex shader? They probably would have mentioned it if it were slower in the high poly tests, so it looks like T&L is ok, but what about the vertex shader? Why might it be slower than the ti4200? Software emulation?

 

[Pete]

DX9 Radeons seem to have unusually powerful vertex shaders. Considering NV30 has fewer VS units than R300 (3 to 4) but a clock speed boost to compensate and still loses, I'd expect a 5600U with 1 or 2 VS units to be much slower than a 9500P's 4 VS units, and probably slower than a 9600P's two.

Let's not get into the 5200.

 

[Slides]

What's a zixel?

 

[pcchen]

I think it's better to describe "zixel" as "Z/stencil write units."

 

[Dave H]

Shouldn't that be a zixel is "the z/stencil value associated with a particular fragment"?

Just as a texel is "the texture value associated with a particular fragment"?

It's hard to get the definition exactly right, but it definitely refers to fragments, not to execution units. (In particular, NV30 "does 8 zixels per clock"; it does not "have 8 zixels".)

In any case...I know demalion will never speak to me again after this, but I formally propose "shixel" in place of his "proxel". Yes, I realize "shixel" sounds like a vaguely derogatory Yiddush term. But "proxel" sounds like a prescription medication for adult incontinence. In my opinion, of course.