Post your breaking NV30 news links here!

Alex, it is only capable of taking one texture sample per pixel pipeline, whatever way you look at it - it can handle 16 textures, but this is 16 texture per pass. I even specifically asked if this was the case in both FP16 and FP32 modes and according to the guy I was talking to he said yes. He said it himself: 8x1.

[edit]: To be more precise "one texture sample per pixel pipeline" shoulkd read "reading from one texture map"

 

at least TomsHardware was very clear on the question of texture samples per pipe:
- with single texture it can do 8 pixels per clock.
- with dual texturing it does four pixels per clock.
- with quad texturing it does 2 pixels per clock.
...and so on.
(so it either has loop back on each pipe, or then it has texel combiner on the end of the pipelines. (btw, is it the latter one even possible with pixel shaders?)

IMO, this is 100% what Dave Baumann has been saying too.

EDIT: got mixed up TMU and Texture Samples...

 

A very simple explanation would be that actually ATi's engineers are so good that they've managed to reduce the transistor requirements by 10%

On a more serious note, the extended shader capabilities must require a lot more registers -> a lot more transistors...


I'm interested about what Carmack has to say about the card.
Also, there are a few things regarding the NV30 and Doom3 that I consider to be questionable - quoting Carmack's year-old statements, supplying Anand with Doom3 performance details... I wonder if it has to do with the leak of the alpha version though.

 

Mixed multi/supersampling.

Well, that AA is really disappointing. If it's all based on GF3/4 modes, 8x won't even beat R9700's 6x qualitywise. Gamma correction is of course a good thing.

 

All preview's I've read so far say an 8x1 design...

 

The exact same situation with the R300. When you are using shaders to combine textures, having more than one texture unit per pipe just wastes resources. No point in having texels available more quickly than the shader can make use of them.

 

Umm, wasn't the .13micron move supposed to reduce heat? If so why the elaborate cooling is needed? What sort of clocking could be managed on the core without such extravagant cooling? Looks like nvidia can't seem to make a cool running core here.

 

On a related note....

Once the NV30 samples start shipping, a most interesting comparison would be:

take a Radeon 9500 Pro and the NV30, and adjust clocks to match. (Upclock the 9500 Pro, downclock the NV30).

This would be quite an apples-to-apples comparison to try and gauge "effective memory bandwidth" utilization between the ATI's and nVidia's architectures: both 8x1 architectures on a 128 bit bus...

 

An extremely nice piece of technology, but considering it'll hit the market almost 8 months after the R300, is the GeforceFX really all that impressive? I really expected something a little more.....spectacular.

All I can say is that I hope by the time boards hit retail, Nvidia better have lower-cost, mass market versions ready to roll...

 

I'm pretty sure that we're talking about the standard 'loop back' as in applying in one pass.

I'm not confident that Tomshardware understood this correct in all cases because, as you point out, it wouldn't make sense with pixel shaders: How would you divide a shader program that depend on several data inputs (texels) over more that one pipeline? You would have to write the intermediate result to a register from one shader, then read it into another shader to carry on. It wouldn't make sense performancewise IMO.

 

[fanboish statement]
The cooling solution shows how hard pressed nVidia was to come up with something to beat the R300. If they couldn't run it at 500 MHz, they couldn't beat the R300 vertex transform rate.

I'm not surprised they targeted 500 MHz in design for a .13 process. But perhaps in design they thought they could get it to run at 500 MHz with lower voltages than it turned out to require. Rather than ship a part with a less convincing lead over the R300, they attach this monstrous cooling solution.
[/fanboish statement]

 

OK, I mentioned this before, but maybe I'm just missing some info that everyone else has or am misunderstanding the quote.

Has Lars just got something wrong (again) or is this just something other previews either missed or didn't emphasize?

 

Why is it a "fanboish" statement to make an observation with regards to stock cooling such as the GeforceFX carries? Obviously the core is running hot.... Anyhow I thought that the .13 micron process was supposed to reduce heat. Clearly it is hot running core if indeed this sort of cooling is needed. It is purely an observation.

 

No, the problem is that nVida's "promise" is simply completely ambiguous:

I mean, what constitutes "abysmal levels" according to nVidia? And what resolutions are we talking about...and on what games...etc. It's completely ambiguous.

It's likely that if nVidia distributed that statement to other reviewers, they just chose not re-state it because it doesn't tell us anything....

EDIT: Or are you confused about gamma correction statements and relative quality?

No one here has disputed NV30's (or Lar's) claim of gamma correction. The comment about NV30 not having improved quality is in reference to the sample grid pattern: the apparent fact that NV30 8X is ordered grid (like GeForce4 4X), and not at least rotated or some other type of custom or programmable grid (like R-300).

 

.13u can reduce power consumed by the part, or not depending on how fast you want to run it and how the part was designed.

.13u can run as low as 1.0V, but maybe timing didn't quite work out at that voltage. If that is the case, they can fix that by bumping up the core voltage to 1.3V or whatever. That will cause the part to use nearly twice the power for the same frequency, but get it to work. (or they could just run it slower at the better power)

 

The review is filled with phrases like that...I'm sorry, I filter them out so I can see why my quote is confusing, but my comments weren't directed at them. I was referring purely to the mention of gamma correction in conjunction with AA and people's disappointment with the AA modes. I'm simply saying gamma corrected AA is a Good Thing and am confused as to why they are disappointed. My only guess as to a reason is the possible lack of Rotated Grid sampling, but it seems to me possible that since other previews don't mention the gamma correction that some aren't including it in their evaluation.

Ah, but the gamma correction mention, if accurate, DOES tell us something not mentioned in other previews.

EDIT: Oh, I'm a bit slow. But the thing is that I didn't see it mentioned in other previews, hence why I bring it up. I consider it an important thing to mention.

 

Bilinear? Trilinear? 'Fast trilinear'?

 

I agree, a confirmation of gamma-corrected AA would be nice to have!

 

http://www.nvidia.com/content/areyouready/twimtbp.html

 

That's the year-old comment I was refering to - thank you for providing me the link to it.

Carmack hasn't updated his .plan yet, though...