[Dave Baumann]

First announced nearly a month back, NVIDIA are readying the GeForce 6600 series for release and now they are ready for the performance numbers to be unveiled for the higher end version, the GeForce 6600 GT. Here we take a full rundown of performance looking at games performance, Shader Performance from Shader 1.1 to 3.0, PCI Express throughputs, video decoding performance, amongst others.

"Looking at the top level specifications it appears that the GeForce 6600 GT bears many similarities to the ill fated GeForce FX 5800 Ultra, however naturally the architecture its based on readdresses many of the performance issues associated with the FX architecture, and the smaller process alleviates the board level issues of noise and heat. Its interesting to note that the specifications for what were considered high end 18-22 months ago are now reaching mainstream price segments."

Read the full review here

[Tim Murray]

the Doom 3 antialiasing comparison is mislabeled (both shots are listed as 6800GT).

PS--wasn't a huge part of the 5800 Ultra's heat problem because of the first-gen DDR2? (well, first-gen DDR2 being overclocked?)

[digitalwanderer]

And in the drop-down box you have "pertformance tests", but typos aside it was an outstanding read.

I liked the Source engine test, and I really like how this card looks. A good mid-level card and the SLI possibilities do tickle the back of me brain.

[pcchen]

In the last paragraph of page 5, there is a "129-bit bus" that should be 128-bit

Anyway, this is a very good preview. I didn't know NV40 can do only 8 pixels per cycle when blending before. Actually I just did a fillrate test yesterday and puzzled by the result as I can get only half of the peak fillrate, but in Ztest it gives near peak fillrate. Now I know where the problem is

[Dave Baumann]

Didn't you know this was a new bus size? :P

Been playing "spot the pipeline counts" with other reviews - looks like very few (bar Hexus, and I have a suspicion why! ) have picked up on the interesting pipeline structure.

[digitalwanderer]

Originally Posted by DaveBaumann

(bar Hexus, and I have a suspicion why! )

*COUGH-COUGH*andy*COUGH-COUGH*

'Scuse me, I got a slight headcold.

[MuFu]

Nice review, as always.

With reference to Anand's results, I'm amazed at how well it does with FSAA enabled in comparison to the 9800XT (i.e. keeps up or beats).

[Rys]

Originally Posted by digitalwanderer

*COUGH-COUGH*andy*COUGH-COUGH*

'Scuse me, I got a slight headcold.

Not quite

We chatted about it over VIA's plans to release an SLI SM4.0 part on 90nm using ATI tech, the other day.

Rys

[digitalwanderer]

Originally Posted by Rys

Not quite

We chatted about it over VIA's plans to release an SLI SM4.0 part on 90nm using ATI tech, the other day.

Doh!

Thanks for setting me straight.

I went and read the Hexus review and really liked it also, they really pointed out a few things that most other reviews ignored/glossed over.

[Tim]

So the NV43 has four "extreme" pipelines.

Still a nice chip as long it has the shader performance of 8 pipe chip, it does not really matter at all.

Edit:

The D3 scores are suprising, it looks like D3 is not fillrate or Z-rate limited it is shader limited.

[Diplo]

Great review, Dave - I was looking forward to reading this one Well done picking up on the pipelines - interesting! Would anyone hazard a guess if the 'problems' effecting Splinter Cell might also spill over into other adapted Unreal Engine games like DX:IW and Thief 3?

Would it be right to assume that the forthcoming AGP version is going to perform almost indentically?

Any idea what price these cards might be retail in the UK? I'm guessing around the £175-£200 mark. If that's the case, they seem to be competiting directly with the 6800 NU. Any chance of adding 6800 NU benchmarks into the fray so we can see how the 6600 GT stacks up against the 6800 NU? If not, which do you reckon would perform best?

[Tim Murray]

Would it be right to assume that the forthcoming AGP version is going to perform almost indentically?

well, yeah, except in the bus benchmarks.

[Rys]

Originally Posted by Diplo

Great review, Dave - I was looking forward to reading this one Well done picking up on the pipelines - interesting! Would anyone hazard a guess if the 'problems' effecting Splinter Cell might also spill over into other adapted Unreal Engine games like DX:IW and Thief 3?

Would it be right to assume that the forthcoming AGP version is going to perform almost indentically?

Any idea what price these cards might be retail in the UK? I'm guessing around the £175-£200 mark. If that's the case, they seem to be competiting directly with the 6800 NU. Any chance of adding 6800 NU benchmarks into the fray so we can see how the 6600 GT stacks up against the 6800 NU? If not, which do you reckon would perform best?

I'm not entirely sure NVIDIA are ever going to release the plain three quad NV40 on PCI Express, so I guess it depends how sites do their comparisons. I know that as soon as I get the AGP version I'll drop the plain NV40 into the article.

Some really basic GPU-bound benchmarks that are bus and platform agnostic do show it to be an interesting comparison though. Here's ShaderMark 2.0, for shits and giggles.

http://img.hexus.net/v2/graphics_car...rk_vs_6800.png

Note that 65.76 breaks a shader test again, compared to 61.34 (and 61.76/61.77 too iirc). Saying that, I really need to find some SM3.0 tests for next time

Rys

[Tim Murray]

isn't the 6800 clocked at 325Mhz? so, wouldn't the clockspeed difference make up for the reduced pipelines? (I was looking for a fraction to use here. then I looked some more. finally, my brain exploded. I guess it'd be 2/3, but whatever.)

then again, the fact that it can operate on 8 pixels at once but can only output 4 at any given time suggests that the 6800NU would have the lead with shaders that aren't very complex (e.g., can be outputted in just a few cycles). OR SOMETHING LIKE THAT.

or maybe they're just detecting ShaderMark with 65.whatever. hehe, had to say it.

[digitalwanderer]

Originally Posted by The Baron

maybe they're just detecting ShaderMark with 65.whatever. hehe, had to say it.

Another typical ATi fanboy comment.

[PatrickL]

Originally Posted by Rys

We chatted about it over VIA's plans to release an SLI SM4.0 part on 90nm using ATI tech, the other day.

My brain is slow but does that mean that via will produce a chipset sli capable and with and ATI SM4.0 graphic core ? In 2006 ?

[Xmas]

The 6600GT is actually ~3% faster regarding shading speed than the 6800. But the latter has a massive bandwidth(+40%)/fillrate(+160%) advantage. So you could say the 6600GT is the 2xAA card, while the 6800 is the 4xAA card.

But thinking about it, the early NVidia roadmaps showing NV41 at the same speed as NV40 make sense. A NV41 at 500MHz wouldn't be a PCIE replacement for the plain 6800, but actually for the 6800GT. And NV45 widens the gap between GT and Ultra a bit.

[Neeyik]

A query about the 4x2 label. The 4 is easy enough to follow but what about the 2? Can the separate pipeline quads operate on the same quad of pixels? The fillrate results in the preview suggest that this is the case because how else would you get 2 texels per pixel per cycle.

[Xmas]

Originally Posted by Neeyik

A query about the 4x2 label. The 4 is easy enough to follow but what about the 2? Can the separate pipeline quads operate on the same quad of pixels? The fillrate results in the preview suggest that this is the case because how else would you get 2 texels per pixel per cycle.

It is really an output limited 8x1 design. That is, it has two independent quad pipelines, writing their outputs to a quad fifo, but if the ROPs can't fetch quads fast enough the pipelines have to wait.

[LeGreg]

Originally Posted by Rys

Note that 65.76 breaks a shader test again, compared to 61.34 (and 61.76/61.77 too iirc). Saying that, I really need to find some SM3.0 tests for next time Rys

I heard that Tommti had a shadermark 2.1 in preparation.
http://www.tommti-systems.de/start.html

[AlStrong]

Originally Posted by Xmas

It is really an output limited 8x1 design. That is, it has two independent quad pipelines, writing their outputs to a quad fifo, but if the ROPs can't fetch quads fast enough the pipelines have to wait.

Why design it like this? Why not something like the R300

[Dave Baumann]

Originally Posted by Neeyik

A query about the 4x2 label. The 4 is easy enough to follow but what about the 2? Can the separate pipeline quads operate on the same quad of pixels? The fillrate results in the preview suggest that this is the case because how else would you get 2 texels per pixel per cycle.

Nothing in current terms really fits what's going on here. However, in terms of peak operations it's peak fill-rate is 4 pixels, and it can achieve that peak when two biliear textures are require, or when trilinear filtering is used.

Realistically, in the fragement pipeline it is more flexible than what you would traditionally associate with an Xx2 pipeline arrangement. However, when looking at peak rates we are looking at pure fill-rate, which it the number of pixels it can write, which is down to the ROP's - in traditional fixed function pipeline terms, the ROP's and the texture units were really all there were.

The definitions are going to get even more meaningless - lets take the theoretical case of a chip that has 8 ROP's, but a unified shader pipeline with 48 ALU's that can be dynamically assinged to different operations; what's its configuration? I'd say it's fill-rate (in traditional terms) is discated by the ROP's.

Originally Posted by LeGreg

I heard that Tommti had a shadermark 2.1 in preparation.
http://www.tommti-systems.de/start.html

Yes, I've been using a beta of that.

[Dave Baumann]

Originally Posted by Alstrong

Why design it like this? Why not something like the R300

Not much point. With a 128-bit bus, even with compression, you are not likely to output much more than 4 pixels in real terms, so you may as well save some transistors on the output whilst maintain a high shader performance.

[jvd]

looks like a pretty nice chip. Wonder how it compares to the 6800nu .

Anyway glad to see it replaxing the 5900 series . Good steps foward .

[Tokelil]

Really nice review.
The NV43 is an interresting chip that seems be less hampered the 128bit interface than I would have though. Seems like a clever idea to save half the ROPs. Can't wait to see ATI's responds.

Anyway, the Texbench (Downstream) benchmark on page 21 is kind of odd. Is the difference between 6800GT and 5900 XT at highest resolution due to HIS or the different plantforms? (AMD vs. Intel)