R520: 32 PS pipelines? Any truth to this?

24 dies per wafer maybe? 1 in 4..

 

But don't forget that most math instructions are single cycle while most texture fetches are not.

 

Regarding yields... I remember someone saying that the price for one processed wafer is somewhere around $3000, but this also depends on process tech and wafer size. Probably 12 inch wafers using a 90nm process cost a bit more than that.

12 inch wafers have an area of about 73000 mm². Considering the size of R430 and G70, R520 might be somewhere in between, maybe 280 mm². Maybe 5% of wafer space is lost due to its circular shape.
That gives a rough estimate of 250 dice per wafer, at a price of $12 each, whether working or not. If they get 10% working, regardless how many pipelines, that's an average price of $120 per die. But that's without packaging and fixed production costs (masks, test runs, etc.). Sounds pretty high to me.

I wonder what kind of yields companies target with such complex chips. Seems like they could make a good profit at 20% ($60), and go dancing in the streets at 30% ($40).

The thought of so much silicon going to waste is somewhat disturbing. Imagine other industries having such low yields

 

I don't forget, there is ever a better metric

 

I've never seen a reliable breakdown of costs on high-end cards at release. The only analysis I've seen made are once they've moved into the midrange and value segment.

My impression is that $60/each to physically produce would not result in a "good profit" for a high-end GPU (which presumably the retail board is selling at $499!), but I'd certainly like to be educated.

Anyone else care to take a crack at it?

 

Well, consider if you will that the manufacturing differences between the very low-end and the very high-end cannot be that much (packaging, board costs). The major differences will be in the RAM and the cost of the chip. So, if you consider that a low-end board sells for $50 or so, I don't see why $150 per chip wouldn't still be profitable for the highest-end boards (though I suppose it does depend on the cost of the RAM).

 

$150/per to produce, Chal. . .or $150/per sales price to the AIB?

I'm also guessing most of the corporate R&D gets assigned to the high-end boards. . .

 

Well, you did get the metric on the sales figures of the fx5700 and 9600xt right? those midrange chips sell for about $75 (at least they did) to aib partners.

I don't know where I read it but the last time I heard someone mention yields on the r420, it was something like 24%.. I'd be happy to figure out who and where said it, but I honestly can't remember..

 

Geo, are you implying that Nvidia is somehow behind the 32-pipeline R520 rumors?

 

Bullshit.

 

I wouldn't necessarily say that Damien - do you personally class NV30 as 4 or 8 pipes?

 

4 pixel shader pipelines.

I don't consider a Z data as a pixel.


While I agree that the number of pipelines alone is meaningless, saying that there are no pipelines is nothing more than bullshit.

 

While it truly is more important what comes out at the other end, NV30 remains a single quad design.

Under that sense it's not off base to claim that R520 is a 4 quad and G70 a 6 quad design, and there's a perfectly good reason also why ATI has to pump up the frequency to new record heights in order to compete. If each R520 quad would be that much more efficient, way below 500MHz core frequencies would had been adequate.

 

Ahh, but its not - there are parts of that pipeline that can handle two quads, yet becuase of that structure I would still consider it to be 4 pipelines (actually, a single quad pipeline) that happens to handle multiple quads in parts of it.

 

From his last post, he looks more on likely to be saying that VR Zone and the like are spinning the story that R520 had 16 pp because it didnt get good yeilds.

 

Same would be true if I take a hypothetical 2 quad design, where each quad can handle more than one quad under conditionals. We have to have at least some baselines to define some things. It'll never be entirely correct with today's complexity.

---------------------------------------------------

As for different websites now claiming that ATI is dumping down quads, that's absolute BS. According to my understanding any desktop R5xx high end GPU were planned with 4 quads in mind since day one.

 

I'm stating that the tea leaves read that way to me. I am not claiming first-hand knowledge. I don't have any "Hey, Psst!" PMs from Brian Burke or Derek Perez on the matter (or any other). Leaving out Inq, an equal-opportunity rumor-monger, the most detailed on 24/32 R520 woes (as opposed to *just* general R520 woes) has come from sites that look to me to have their best sources in the NV camp day-in/day-out (whether NV themself, or their AIBs, or whatever). Further, if you look around, you might find a post or two hinting in that direction from what I fondly refer to as "our wise guys" (finger beside the nose "wise guys", as opposed to Gold, Frankincense & Myrrh). I won't embarrass any of them by specifically pointing at them.

But I could be wrong. Wouldn't be the first time.

Of course the hugely amusing thing would be if they are right, and we are wrong. The VR Zone post cited above, however, looks like clear evidence to me of a rapid retreat on that front, spewing smoke to obscure it all the way.

 

I shudder to think about the mental state of those who don't have the warm embrace of B3D to shelter them from the mad world of extremes out in the wilderness.....

 

It doesnt make sense for me for Nvidia to hype the R520 over the short term period when the G70 has zero competition. Even if Nvidia was trying to hurt peoples expectations of the R520 it could potentially damage the sales for the people who standing on the line of uprgrading and the "Wait and see" crowd. Even if the R520 arrive and doesnt turn out to be a killer. It'll still be a competing product with the G70. I have feeling Nvidia wants to capital on the fact the Geforce 7800GTX is alone right now.

 

And over the last month or so, R520 has been working its way back to "a failed 32-pipe card released at a crippled 16-pipes". Is there any other way to read the VR Zone piece? And Josh's piece two weeks before 7800GTX was released?

Of course, I have a touchstone, and if the touchstone is wrong then I'm way the hell out in left field --the touchstone is R520 is, and has always been, a 16-pipe card.