NVIDIA GF100 & Friends speculation

[AlexV]

I compare the pictures of G80, GT200 and GF100. And i saw this:

....

I think GF100 has a g80 die size.

Ugh, would you mind detailing the approach you used in your comparison? What did you choose as SR in order to ensure that you're actually comparing comparable things, how do you know the package:die ratio and so on and so forth. I'm certain you took all those precautions to arrive at such a decisive conclusion, no?

 

[aaronspink]

A quick pixel counting exercise makes it ~500mm/sq, using the width of the longest (second) section of PCIe x16 connector as a reference. I make that 71.65mm according to wikipedia, so ~0.746mm/pixel (96 pixels wide), and I make the die area to be 30x30 pixels.

That's a five minute go for cheap thrills anyway

Why pixel count when we have a picture with a measuring stick on it. The package itself is just over 4cm per side. Assuming .75cm side margins, the die could be as big as 25x25mm.

 

[aaronspink]

I compare the pictures of G80, GT200 and GF100. And i saw this:

http://www.xtremesystems.org/forums/showpost.php?p=4268014&postcount=1651

I think GF100 has a g80 die size.

You can't tell anything except approximate maximum die side given a measurement of the package/ihs. Side margins can range from .25 to .75 cm depending on the technology. All this does is put an upper bound on the die size, and that is ~625mm^2. The package size is similar to the G80 only because they both have approximately the same number of pins. Given the G80/G100 package size, you could reasonably fix a 900mm^2 die under it if such a thing was possible if one used .5cm side margins.

 

[fellix]

Here, I've made a more precise alignment. The overlayed semi-transparent image is a GTX280 board:

For scaling reference, I used the SLi fingers and ATX power connectors.

This is naked GTX280 (576 mm² 65nm die):

As you see from the picture, it is definitely possible to fit ~550mm² die on the smaller GF100 organic substrate.

 

[Rys]

Why pixel count when we have a picture with a measuring stick on it.

Because you can see the die area delineated in the package area in the CAD drawing.

 

[Rys]

Unless of course that gap in the balls means nothing

 

[neliz]

Unless of course that gap in the balls means nothing

Are we talking about bifurcated... wait.. Rys' doublepost?

Depends on which of the two pics is real.. right?

The tweakers.net picture, with the ruler next to it.. what's up with the black imprint on the package? couldn't that mean something? it's awfully close to 25mm on the ruler..

here is the pic with the ruler: http://www.heise.de/imgs/18/4/8/9/6/3/1/3.jpg-90389ba714eca01e.jpeg
here is the pic that shows some kind of imprint/package dents: http://www.heise.de/imgs/18/4/8/9/6/3/1/1.jpg-e3443478528b80c7.jpeg

 

[Thraxis]

Might be something completely different but the spacing in the chip is due to the pinout correct? Not related to a complete die shot. Or is it the other way around?

Courtesy of geeks3d. http://www.geeks3d.com/public/jegx/200910/nvidia-gt300-gpu-back-face.jpg

 

[Alexko]

Are we talking about bifurcated... wait.. Rys' doublepost?

Depends on which of the two pics is real.. right?

The tweakers.net picture, with the ruler next to it.. what's up with the black imprint on the package? couldn't that mean something? it's awfully close to 25mm on the ruler..

here is the pic with the ruler: http://www.heise.de/imgs/18/4/8/9/6/3/1/3.jpg-90389ba714eca01e.jpeg
here is the pic that shows some kind of imprint/package dents: http://www.heise.de/imgs/18/4/8/9/6/3/1/1.jpg-e3443478528b80c7.jpeg

Notice in the first picture how the guy blurred the markings on the packaging, but not his fingerprint? I doubt NV would go that far to find the leak, but it's still amusing.

 

[neliz]

Notice in the first picture how the guy blurred the markings on the packaging, but not his fingerprint? I doubt NV would go that far to find the leak, but it's still amusing.

In the article they said that NV didn't allow partners to show cards, but they had no problem making shots and they weren't bound by NDA despite having a visit from them last week, because the marketing guy visiting them wasn't allowed to talk about GF100..

 

[TKK]

So going on that, and taking THG's benchmarks - http://www.tomshardware.com/reviews/radeon-hd-5970,2474-8.html -

The 470 beats crossfired 5870's without AA, but loses to the 5850 with 8xAA.

This seems a tad unlikely but if true at least we know what kind of benchmarks to expect come March 26th.

Unlikely but not impossible. We don't know how much the 8xAA efficiency was improved over GT200. If there wasn't much of an improvement in efficiency, the GTX 470 is bound to show similar 8xAA performance hits to the GT200.

As XMAN26 wrote (too lazy to search for his post now), the remaining parts of what was once called the core, the ROPs and L2 cache, are probably clocked at 450-475 MHz. The GTX 470 has 8 ROPs disabled, so:
475*40=19000

compared to GTX 280:
600*32=19200

The GTX 280 also has 10% more memory bandwidth.


What could also have affected the scores is AF performance. While Nvidia claims significant efficiency gains over GT200, we have to keep in mind that for the GTX 470, it's 56 TMUs @ 625 MHz vs. 80 TMUs @ 650 MHz on the GTX 285.

In summary, the GTX 470 has less pixel fillrate, less memory bandwidth, and (considerably) less texture fillrate compared to a GTX 280/285. Under these circumstances, a stronger performance hit at higher levels of AA/AF is not all that unlikely.
And yes, I'm entirely aware that this is pure guesswork for now and that paper specs don't necessarily reflect real world performance, but it's one possible explanation why the GTX 470 might suffer that much with AA/AF maxed out.

 

[LonelyMind]

Unless of course that gap in the balls means nothing

Only if they get 'em kicked hard, performance wise.

 

[MfA]

As long as someone is making bad jokes about that statement ... guess we know it's not gay.

 

[DegustatoR]

A quick pixel counting exercise makes it ~500mm/sq, using the width of the longest (second) section of PCIe x16 connector as a reference.

Which puts it in G80 and GT200b area. Still big but not "huge".
And that would mean that NV maintained it's top end die size between generations while AMD went with a bigger Cypress.

 

[Mindfury]

The Chinese name card is most likely GTX470,but I'm not sure.

Another rumor from Chinese forum:
http://we.pcinlife.com/thread-1366899-1-1.html

 

[Thraxis]

~500mm2 would need ~50% increase over the 5870 to compete for Performance/mm2 however. Just a stray thought that entered the mind.

As XMAN26 wrote (too lazy to search for his post now), the remaining parts of what was once called the core, the ROPs and L2 cache, are probably clocked at 450-475 MHz. The GTX 470 has 8 ROPs disabled, so:
475*40=19000

Doubting it will be clocked that low, especially with the supposed leak of the Chinese benchmark showing a pretty visible 625 (If true).

 

[Thraxis]

(As I was sniped with a benchmark in Chinese and still don't have edit button, talking of a different benchmark)

Shows the 448 at 625 and 1250/1296 clocks and doubting the 480 will have such a great dip in clock speeds if it wants to even compete against it's cut sibling.

 

[digitalwanderer]

The Chinese name card is most likely GTX470,but I'm not sure.

Another rumor from Chinese forum:
http://we.pcinlife.com/thread-1366899-1-1.html

Is it just me or do those numbers almost look like they may just possibly be real?

 

[GZ007]

And that would mean that NV maintained it's top end die size between generations while AMD went with a bigger Cypress.

Thats a great achievement from NV .This could be used by Nvidia PR

 

[Sontin]

(As I was sniped with a benchmark in Chinese and still don't have edit button, talking of a different benchmark)

Shows the 448 at 625 and 1250/1296 clocks and doubting the 480 will have such a great dip in clock speeds if it wants to even compete against it's cut sibling.

It's 1296Mhz - also 625Mhz ROPs / 648Mhz TMUs, Setup... / 1296MHz Hotclock.