G70, G71, R520, R580 die size discrepancy

[atomt]

Beyond3D list the folowing as dimensions (from reviews )
G70 - 334mm^2
R520 - 288mm^2
R580 - 352mm^R2
V530 - 156mm^2

Techreport
http://techreport.com/reviews/2006q1/radeon-x1900/index.x?pg=1
G70 - 333mm^2
R520 - 263mm^2
R580 - 315mm^2
RV530 - 132mm^2

Nvidia Slides
http://www.nvnews.net/vbulletin/showthread.php?t=65835
G71 -196mm^2
G73 - 125mm^2
R580 - 352mm^2
RV530 - 149mm^2

I estimated the G71 die size to be ~ 250mm^2. I ASSUMED that the heatsink
mounting holes are the same for 7900GT and 7800GT boards. I measured the die dimension and scale it using a G70/334mm^2 as a reference.

From this pictures, where what appears to be a G73 on a 7900GT board,
I estimate the G73 to be ~ 150mm^2. As a reference , Samsung GDDR3
FBGA package is 154mm^2 (from Samsung specs)
http://www.nvnews.net/vbulletin/showthread.php?t=65257

Beyond3D measurements for Nvidia die size is same as Techreport. Techreport estimates
for ATI die sizes are smaller than Beyond3D.

A ruler is not accurate but 25% difference?

Nvidia and ATI count transistors differently but it appears that Beyond3D, TechReport
and Nvidia use different rulers as well. Samsung ruler is different from Nvidia as well.

90nm transistor area is 67% that of 110nm. In a shrink, due to ML and routing limitation,
the shrink is < ideal. Even assuming an ideal 67% shrink, G71 die size = 0.67 x 334mm^2
= 224mm^2. Is Nvidia that much better at fitting 300M transitors into 196mm^2 while ATI
needs 288mm^2 for R520 321M?

 

[Arun]

R580 - 352mm^R2
V530 - 156mm^2

R580 - 315mm^2
RV530 - 132mm^2

R580 - 352mm^2
RV530 - 149mm^2

Obviously, the middle figures are fucked. Sounds like a 1mm error on both dimesions to me, or being lazy and linearly scaling based on NVIDIA's density...

From this pictures, where what appears to be a G73 on a 7900GT board

I'd beware personally - you don't know if it's the same revision etc.

Is Nvidia that much better at fitting 300M transitors into 196mm^2

At hinted towards in the other G71 thread, G71's transistor count is below that of G70. NVIDIA's transistor density IS higher, though.


Uttar

 

[Sunrise]

Beyond3D measurements for Nvidia die size is same as Techreport. Techreport estimates
for ATI die sizes are smaller than Beyond3D.

A ruler is not accurate but 25% difference?

Nvidia and ATI count transistors differently but it appears that Beyond3D, TechReport
and Nvidia use different rulers as well. Samsung ruler is different from Nvidia as well.

I just looked at this myself and i came to the following conclusion:

It looks like NV compares physical die size of R580 with their die size for G71 from the data sheets (before the die is cut from the wafer) which is as wrong as it gets when trying to compare both cores.

You should ignore that altogether, as this is quite missleading.

Let´s wait for Dave´s or Degustator´s measurements of an actual G71 core, for further judgements.

 

[atomt]

Obviously, the middle figures are fucked. Sounds like a 1mm error on both dimesions to me, or being lazy and linearly scaling based on NVIDIA's density...

I'd beware personally - you don't know if it's the same revision etc.

At hinted towards in the other G71 thread, G71's transistor count is below that of G70. NVIDIA's transistor density IS higher, though.


Uttar

That picture ain't a G71, so unless Nvidia have a 160mmm^2 chip. The PCI-E signals are not even connected to the package.

G71 transistor count need to be 10% less than a G71 to shrink from G70 334mm^2 to
196mm^2 assuming an ideal 67% shrink. So G70 have 10% useless transistors?

 

[Arun]

That picture ain't a G71, so unless Nvidia have a 160mmm^2 chip.

I suspect a NV43 (6600GT) chip, personally:
http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=71
http://www.beyond3d.com/forum/../misc/chipcomp/?view=chipdetails&id=7

G71 transistor count need to be 10% less than a G71 to shrink from G70 334mm^2 to 196mm^2 assuming an ideal 67% shrink. So G70 have 10% useless transistors?

Not sure, but if you look at past generations, it doesn't feel impossible (although G70 wasn't really a "next generation" chip.)

Uttar

 

[atomt]

I suspect a NV43 (6600GT) chip, personally:
http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=71
Uttar

The chipcaps location is different. The package is identical down to all chip caps
position to G73 pictures in vr-zone and others. The funny thing is that the card
have 8 Samsung 32-bit GDDR3, so 256-bit memory interface. On top of that,
the PCI-E signals are not even connected. Note the 2 rows of unused solder pads all around the chip package. Dailytech have since removed the card picture but this site have
the original dailytech pics. Since that chip is smaller than 196mm^2, so it ain't
a G71, what is that? A G73 with 128-bit memory interface on a 7900GT board?
The package pins out are different.

http://publish.it168.com/2006/0221/20060221256501.shtml

 

[atomt]

Die saw street

I just looked at this myself and i came to the following conclusion:

It looks like NV compares physical die size of R580 with their die size for G71 from the data sheets (before the die is cut from the wafer) which is as wrong as it gets when trying to compare both cores.

You should ignore that altogether, as this is quite missleading.

Let´s wait for Dave´s or Degustator´s measurements of an actual G71 core, for further judgements.

Die saw street is 80um for 90nm low-k. So not significant.

We just have to wait until some get hold of a board and use the "correct" ruler.

 

[atomt]

Obviously, the middle figures are fucked. Sounds like a 1mm error on both dimesions to me, or being lazy and linearly scaling based on NVIDIA's density...
Uttar

TechReport ATI die sizes are dffferent from Nvidia and Beyond3D but their
Nvidia chips measurements are close to Beyond3D. TechReport have 2 rulers,
one for ATI and one for Nvidia? ATI R580 does look pretty big. I am inclined to
trust Beyond3D since Dave measure the die size for almost every card. Other
sites don't do it consistently. Most of the time, no die measurements.

 

[rwolf]

Don't forget that ATI chips have built in support for GDDR4 and virtual card memory. This takes a fair number of transistors.

 

[Sunrise]

Die saw street is 80um for 90nm low-k. So not significant.

Certainly not that significant, but there´s still a disparity between those sizes, then.

 

[Sunrise]

I am inclined to trust Beyond3D since Dave measure the die size for almost every card. Other sites don't do it consistently. Most of the time, no die measurements.

Well, that would be a direct question to Dave. I´m also inclined to believe his 3D tables, but something doesn´t make sense here.

Dave, if you are out there, can you please confirm that you did those measurements yourself? Not just on R580, but also on G70? I´ve just accidentally stumbled over the die sizes of G70 and R580 respectively, and while G70 looks to be quite big, you have it listed smaller than R580. Is that actually the case ?

 

[Xmas]

It seems to me the B3D figures for R520 and R580 are off. R580 should be smaller than G70.

Not sure, but if you look at past generations, it doesn't feel impossible (although G70 wasn't really a "next generation" chip.)

I've always thought G70 was much too big in relation to NV40 (which wasn't exactly economic regarding transistors either) and had expected around 250-260 million transistors for a 6 quad part before its launch. I guess after good initial results from G70, they thought the transition to 90nm along with a little "cleanup effort" could keep them competitive while shifting all resources to future projects.

Certainly not that significant, but there´s still a disparity between those sizes, then.

Less than 0.5%.

 

[Sunrise]

Less than 0.5%.

Yeah, pretty insignificant, actually. Should´ve worded that differently.

It´s funny though, that no one even questioned those die size measurements until now.

 

[silent_guy]

Less than 0.5%.

<Pendantic mode on>
Hmmm. It's a bit more than 1% actually. Assume a scribe of 80um (aggregate, both sides) for 200m2, then you get:

(sqrt(200mm2)+0.08m)^2 - 200 = 2.3/200 = 1.15%

But are you sure it's 80um for both sides and not 160um? That would increase the difference to 4.5 mm2.
<Pedantic mode off>

 

[atomt]

Saw street/Dicing blade

<Pendantic mode on>
Hmmm. It's a bit more than 1% actually. Assume a scribe of 80um (aggregate, both sides) for 200m2, then you get:

(sqrt(200mm2)+0.08m)^2 - 200 = 2.3/200 = 1.15%

But are you sure it's 80um for both sides and not 160um? That would increase the difference to 4.5 mm2.
<Pedantic mode off>

The saw street is 80um. The cut (kerf width) is always slightly wider than the blade used.
I can't remember the exact percentage achievable since it have been a long time since I worked in dicing but it should be ~ 50%. Assuming a 20um (0.8mil) thickness blade is used, the kerf width is 30um leaving 25um each on both sides.

http://www.flipchip.com/services/die_level/dicing/

Bottom line is that it is still insignficant compared to the discrepancy in die sizes.

 

[Razor1]

Was the original die size of the g70 a bit bigger then expected?

 

[atomt]

Was the original die size of the g70 a bit bigger then expected?

Assuming B3D measurements of G70 is correct (334mm^2), and Nvidia R580 slide of R580 being 352mm^2 is the same as B3D measurements, the G71 die size claimed by Nvidia slides seem too small (it appears like either 186mm^2 or 196mm^2 in the slide)

I measured the G71 die size in the picture. Since there is no scale, I used the heatsink mounting holes as a reference and compare that with a G70. I approximated the G71 die size to ~ 250mm^2.

The 2 question I have is :

1. A shrink of 334mm^2 to 196mm^2 is a 58% shrink. A 90nm transistor is 67% the area of 110nm transistor. Shrink efficiency is often lower because the analog/buffer circuits do not shrink as much due to need to maintain drive currents
(interface to PCI-E/GDDR3, DVI, RamDACs)

2. That would make G71, almost 1/2 the size of R580. Is it possible that G71 architecture so much more efficient or R5xx is ATI's GeForce FX?

 

[Jawed]

I was under the impression that TSMC's 110nm is a half-node, meaning that some features are shrunk, but others are not.

If that's so, then it means that 110nm produces a larger die than would be expected.

It would also mean that %age scalings from 130nm to 110nm and from 110nm to 90nm are invalid. Let alone the reasons given earlier for the way that different kinds of features (e.g. RAM) scale differently with varying processes.

So, all in all, it seems like a fools-errand to evaluate die size at 90nm based on 110nm.

Jawed

 

[trinibwoy]

I was under the impression that TSMC's 110nm is a half-node, meaning that some features are shrunk, but others are not.

If that's so, then it means that 110nm produces a larger die than would be expected.

It would also mean that %age scalings from 130nm to 110nm and from 110nm to 90nm are invalid. Let alone the reasons given earlier for the way that different kinds of features (e.g. RAM) scale differently with varying processes.

So, all in all, it seems like a fools-errand to evaluate die size at 90nm based on 110nm.

Jawed

Even so, based on what you've pointed out, it makes the < 200mm^2 claim for G71 even more ridiculous.

 

[fbomber]

Maybe the chip is less complex, transistor count wise (less rops, pipes, or whatever), than the g70, but ends faster because of the clock speed. The real monster would be the 2 chip version.
This would explain some important developer hinting about a 2 chip reference card from a major IHV.