G71/G73: Summary & Pre-Launch Speculation Frenzy

[KimB]

* 200 posts OMFG! *

n00b

 

[no-X]

Maybe ATI should come out with a slide "ALU for ALU, we are 150% more powerful!" or some such nonsense

ATi could say, that G71 (GTX) has 50% more fillrate and 3% more bandwidth than R580 (XTX) and still is slower.

 

[_xxx_]

Configurability is presumably being able to produce at least 3 SKUs based on only one die and two PCBs: GTX, GT, GS.

Jawed

That's how I got that, too.

 

[N00b]

n00b

Someone summoned me? What can I do for you, master?

 

[atomt]

The Inquirer

http://www.theinquirer.net/?article=30100

what makes'em think so !?

The Inquirer article mention use of 6600GT PCB which only use 128-bit memory
interface. 6600GT also only have 4 GDDR3 memory chips.

The Samsung memory chip mentioned is 64MB chip (512MBit device, 32 bit interface)

To get 256-bit, need 8 GDDRs chips which means 512MB of GDDR3.
BC-14 memory, so capable of 1400MHz. (Inquirer says it is overclocked)

7600GT having more memory bandwidth than even 7900GT?

Nah, Just stupid reporting by Fuad.

 

[mikechai]

n00b

Thanks for the comment anyway.
________
UHWH

 

[Maintank]

Edit: NM rereading it, you mention the 7900 Duo in the Misc section.

 

[Coz]

Scary???? Given the fillrate (and MB) advantage that these new parts have over the ATI parts, its not scary at all...

The 7900GTX has got the same pixel fill-rate as the X1900XTX and only 3% more memory bandwidth. It has a 50% advantage in texture fill-rate, but is that really such a big deal in modern games (ie. shader-intensive)?

 

[aaronbond]

Speaking of MSRP - http://www.mwave.com/mwave/DeepSearch.hmx?scriteria=7900+GTX&ALL=y&TP=2

EVGA 512-P2-N570-AX GF7900GTX EGS 512MB PCI-E W/HDTV & DUAL DVI

$765.00

EVGA 512-P2-N575-AX GF7900GTX SUPERCLOCK 512MB PCI-E W/HDTV & DUAL DVI

$795.00

Does HDTV mean it has HDCP support? I really want my next card to have support for it. I understand it's another component, but is the norm going to have it on-board w/ this 79xxx series? What will make my decision even greater is if DX10 is shipping w/ Vista? If so I'll wait on purchasing a new card. I don't understand all these refreshes of basically the same card. Guess they're keeping up w/ the Canadians. =D

 

[Geo]

Geo, Dave said that ATI and NV counted transister differently. It seemed like it was some time ago thus it may or may not be still relevent....

Yeah? Linky, please.

 

[Geo]

Geo, Dave said that ATI and NV counted transister differently. It seemed like it was some time ago thus it may or may not be still relevent....

After consultation with JB in IRC, it turns out this is the post he had in mind:

http://www.beyond3d.com/forum/showpost.php?p=270073&postcount=51

It's from Chalnoth. Chal, where did you get that impression?

Edit: http://www.beyond3d.com/forum/showpost.php?p=246677&postcount=155 And there's Lord Wavey himself. . .

 

[dizietsma]

dizietsma: Geo had made the interesting comment a while back that for a given wafer with a process with roughly similar defect rates, the number of defects over the wafer is constant, yet there are more potential chips on it. As such, the number of chips with defects will be statistically lower. Since G70 was 334mm² and G71 is 196mm²... Well, you can guess the rest.

yes but you will still have some defects .. I therefore conclude you are goingn for another card with 5 quads to soak up the defect solution ?

 

[Farid]

The thread topic is G71/G73: Summary & Pre-Launch Speculation Frenzy, in case some people forgot.

If you want to keep on the discussion about the logic behind the R580 architecture and its emphasize on Shaders, I'll just split the thread so we can continue both discussion in their own threads.

 

[Farid]

Threads Split

Ok, after a little discussion with some folks, it seems that it's indeed better to have a thread for each of the two discussions going on.

So here it is:

G7x vs R580 Architectural Efficiency
http://www.beyond3d.com/forum/showthread.php?t=28973

 

[yacoub]

So the G71s are true G70 refreshes (as they were always intended to be): Smaller die size, higher clocks. This leads to the mediocre 3DMark score increases over 7800 series cards but it should be enough to match the X1900XT and maybe the XTX, when it comes to the 7900GTX. Also with the better pricing paving the way for dual-GPU cards coming to market at more mainstream market prices in the coming months, a 7900GT for $300 makes sense and I look forward to it reaching that price in short order. Also will be interesting to see how the X1900XL competes with it once released around the end of the month, probably in coordination with the ATI price drops on the XT and XTX.

 

[Geo]

There've been increasing hints that X1900xl isn't coming just now. Unless, as some wit pointed out in IRC, you want to get one with a TV tuner (X1900aiw).

 

[compres]

Ok, after a little discussion with some folks, it seems that it's indeed better to have a thread for each of the two discussions going on.

So here it is:

G7x vs R580 Architectural Efficiency
http://www.beyond3d.com/forum/showthread.php?t=28973

Ouch, kind of hard to find my last read post on that thread(tough day for a lurker).

Regarding the 7900gtx, are we really speculating anything? Seems to me the launch is inminent and final info is already known about it(leaked).

 

[Geo]

Ouch, kind of hard to find my last read post on that thread(tough day for a lurker).

Regarding the 7900gtx, are we really speculating anything? Seems to me the launch is inminent and final info is already known about it(leaked).

Well, there was that "2x floating point" thingy from dailytech that doesn't seem to add up to anything else we know. And there's always "that ole' black magic that you weave so well" from the NV driver team to take into consideration. Other than that, nothing is coming to mind. But, y'know, until it's official it ain't official. Would I be really surprised at a major surprise? Sure. Nevertheless.

 

[3dcgi]

Well, by some measure, transistor count is more tightly-connected to the underlying architecture than die size, so it's somewhat nice to have the information for speculation on how architectures will scale over different numbers of pipelines, and how well they'll do at different processes.

But you're absolutely right: for the most part, transistor count is just a proxy for die size in these speculations. It is the die size that is the important bottom line. I think we mostly make use of transistor count because it's largely independent of process technology, and we're used to thinking in that way on these forums.

This thinking might be fine if transistor count was something more than a guess. As silent_guy said engineers don't waste time trying to count transistors. At least none that I've ever worked with or talked to measure anything other than die size.

 

[silent_guy]

[*]"three" is a pretty funny number for a computing architecture. I wouldn't be surprised if each shader unit in R580 (and RV530) is actually composed of four quad-pipelines, with one dropped for the sake of redundancy. That's 25% redundancy, on a total of approximately 128M transistors (64 pipelines, 16 dropped for redundancy). If that's the case, then we prolly won't see a "36 pipeline" variant of R580.[/LIST]Obviously, all guesses... The point being that R580's "size" might look like a huge disadvantage, but there are hierarchies of fine-grained redundancy at play, and I expect that redundancy in R580 is significantly more advanced than R420. I also suspect that the "3:1" architecture of RV530 and R580 adds a significant layer of redundancy. The end-result being, perhaps, that practically every core comes out as a fully functional R580 (or RV530).
Jawed

Hmmm. There's definitely a possibility that there are a number of excess units for redundancy. But an overhead of 33% (because that's how you should look at it!) is out of the question.

Let's use some numbers from the past (I don't have the latest info). It is perfectly possible to get yields of >70% in a 180um process for dies of 100mm2 *without* using any redudancy.
When you think about it, that's really staggering: if you have 300 dies per 8" wafer, that means that you have only around 100 defects per wafer. It's actually a bit more, because multiple defects per chip do not decrease yield in this case.
Looking at the histogram of the amount of defects per die, you'll see that 70% have 0 defects, 15% have 1 defect, 10% have 2 defects etc. (Just making up the numbers, but it's something like that.)

Using the defect/die distribution above, it means that adding just 1 redudant atomic block, will increase the yield from 70% to 85%. 2 redudants blocks increase it from 85% to 95% etc. Clearly, the bang/buck for each additional redundant block does down very quickly.
Now this is an ideal case, where the core of the chip is nothing but exact copies of the same atomic blocks. (If you ever wonder why DRAM's have production yields above 98%, here's your answer.)

In the real, non-DRAM, world, you have different functional blocks. E.g. a number of identical DSPs, transmogrifiers, pixel shaders, what not. If you want redundancy there, you'll have to add redundancy for each of those individual units. Obviously, this complicates matters somewhat, but trust me that all big chip houses have software to calculate the !/$ ratio for adding or not adding redundancy for each of those units.

Defect density for 90nm must be quite a bit higher than 180um. And a die of 350mm2 instead of 100mm2 also increases the number of defects per die so a higher redundancy may be needed, but I'd be very *very* surprised if it's higher than, say, 10%.

Note also that for the chip of the same size, you need less redudancy if you have more but smaller atomic blocks instead of less but larger blocks, since in the former case you have a higher granularity of enabling or disabling a defective unit. The 580, with it's 48 pixel shaders, is probably in that camp.

As for the multiple of 3 instead of a power of 2: nah, that's really not an issue. It means that some 2-bit busses will carry numbers from 0-2 instead of 0-3, which wouldn't make Claude Shannon proud in terms of information code density . , but that's about all there is.