AMD: R7xx Speculation

[Lukfi]

=>chavvdarrr: Charlie Demerjian claimed $100-150 per chip...

 

[Wirmish]

GT200b will support GDDR5 and it is coming sooner than you think.

The early G200 buyers are going to be very happy about this...

 

[BRiT]

R700 is still an unknown quantity at this point. If it does actually have a shared memory architecture, it may do a much better job of AFR than any previous product. If not... well, in games where AFR works well, it should quite significantly outperform GTX280 for about the same money. How many games that is remains to be seen.

If not, it may generate higher numbers, but will not deliver as smooth of a game-play experience. The only way is if the numbers are twice that of a single card, which would require a single HD4870 to be on par with a GT280.

 

[Arty]

not at all, at certian parts of the game yes, but those are all unplayable settings on most of those games on both cards! We already know when the g92's hit a bandwidth/memory bottleneck they take a nose dive, but why use benchmarks that aren't even playable on a HD4950 too? They are pretty much pointless other then quake, doom and farcry.

A fair comparision (price excluded at this point) would be against the 9800 gtx, where bandwidth is similiar. The price of the gtx will be lowered in due time, as will the 9800 gt which will undercut the 4950. nV does have room to play here. .

Thanks. Unplayable fps has not and will not stop either IHV from showing off their leads.

AMD is only taking around 25-30% on margins for their new chips, nV has been taking around 35-40% for thier old chips. So they do have some space for a price war. It all depends on how fast nV ramps up the g92b though, I think its a straight shrink but they can clock these chips higher, since the clocking room for the g92 is very good only limited by voltage.

This has always been the case. However, one thing to remember is that traditionally margins are more dear to Nvidia than AMD ..

 

[Hekking]

Didn't GDDR 3 also sucket pretty much watt's compared to GDDR 4/5?

 

[Arun]

Per unit of bandwidth, yes, GDDR3 definitely takes more power than GDDR4 and especially GDDR5. Per bit, it's a bit more complex, but GDDR5 remains very impressive AFAIK.

 

[Arty]

Per unit of bandwidth, yes, GDDR3 definitely takes more power than GDDR4 and especially GDDR5. Per bit, it's a bit more complex, but GDDR5 remains very impressive AFAIK.

Are there any figures for GDDR5 efficiency?

 

[Geo]

I have never seen so much disinformation on B3D.

What is happening?

I've never seen a member before now with a pink elephant growing out of his ear.

See how easy that assertion thing is? Anybody can play. But it's a very tiresome game.

If you want to talk at length about what rumours that are out and about in the community that you believe to be misinformation about R7xx, and why you think they are misinformation, feel free. Lord knows we are deep into silly season, just before the dawn of real facts.

But unless you are willing to have that conversation on a point by point basis you're just trolling, and we don't like trolls around here.

 

[Wirmish]

MAGIC PC (french)

Club3D - Carte Graphique - ATI - HD4850 - 512 Mo GDDR3 - Dual DVI - HDCP - Dispo lundi 16 juin - 162,00EUR

Super Trad. :

Club3D - Graphic Card - ATI - H4850 - 512 MB GDR3 - Dual DVI - HDCP - Available monday june 16th - 208.18USD

MONDAY JUNE 16 !!!

 

[Lukfi]

Yeah, I've heard about retailers here in the Czech Rep. already having the cards, or having them on the way, ready to sell next week. So it's definitely not a rare occurence you've just found

 

[MarkoIt]

And by that time comes RV770 in 40nm, and them we will see who catch who

40nm is the half node of 45nm witch is a new process node. It decrease dramatically die-size.

Ati RV870

2000sp -> 3 Teraflops ...
64TMU
24Rops
256bit@GDDR5 5 Ghz
250 mm^2@40nm
750 Mhz
Avviable in June 2009
-> GT200b/c/d/e owned






I did some calc.. is it sensible that block TMUs(32) of RV770 take about 60 mm^2 of the die-size?

 

[Wirmish]

Yeah, I've heard about retailers here in the Czech Rep. already having the cards, or having them on the way, ready to sell next week. So it's definitely not a rare occurence you've just found

Yeah... already selling in Hong Kong.
Selling next week in Czech Rep., France, USA (Indiana), ...

ATi's NDA ends june 25 ?
This is madness !!!

 

[Ninjaprime]

Based on some rumors:

Die size: 24mm x 24mm -> 576mm²
Wafer: 300mm
Dies/wafer: 89 ~ 92
Yields: ~40%
Good dies/wafer: ~36

GTX 260 = ~27 dies/wafer
GTX 280 = ~9 dies/wafer

I think you're greatly overstating how many dies on a wafer:

http://www.tomshardware.com/reviews/geforce-8800,1357-2.html

The 8800gtx only put 80 dies on a wafer, and it was only a 484mm². GTX 280 being 19% larger in size, you can therefore conclude they are only getting about 65 dies per wafer. If they are really only getting 40% yields then they are only getting something like 26 good dies per wafer.

Ouch.

 

[Wirmish]

I think you're greatly overstating how many dies on a wafer:

http://www.tomshardware.com/reviews/geforce-8800,1357-2.html

The 8800gtx only put 80 dies on a wafer, and it was only a 484mm². GTX 280 being 19% larger in size, you can therefore conclude they are only getting about 65 dies per wafer. If they are really only getting 40% yields then they are only getting something like 26 good dies per wafer.

Ouch.

But I see 118 dies on this wafer...

ExtremeTech also -> 118 dies


Try with this -> G200 = 24mm x 24mm

 

[Ninjaprime]

But I see 118 dies on this wafer...

ExtremeTech also -> 118 dies


Try with this -> G200 = 24mm x 24mm

Hmm I just did a hand count, seems you're right, Toms must have meant 80 good dies on that wafer... maybe?

So if it is 40% and thus 36 dies per wafer, whats the cost of the wafer? I can't imagine its gone down since the 8800gtx days, so with half or less dies compared to G80 its probably safe to say the per chip cost of GTX 280 is at least twice the cost.

What did 8800gtx cards launch at, $550?

 

[Arun]

Hmm I just did a hand count, seems you're right, Toms must have meant 80 good dies on that wafer... maybe?

That is correct, we were told ~80+ usable dies/wafer at G80's launch. Presumably yields increased a bit over time, so who knows maybe they were at 90 usable/wafer with A13 in mid-2007.

 

[Wirmish]

Die size:
GT200: 576 mm²
RV770: 256 mm² (or 276mm²)

Dies / 300mm wafer:
RV670 (55 nm) = ~292
RV770 (55 nm) = ~221 (~200 if RV770 = 276mm²)
G80 (90 nm) = 118
G92 (65 nm) = ~166
G92b (55 nm) = ~200
GT200 (65 nm) = ~89

Production cost:
GT200: ~$105
RV770: 89 / 221 * $105 = ~$42 (or $47 if RV770 = 276mm²)

 

[Jaaanosik]

Die size:
GT200: 576 mm²
RV770: 256 mm² (or 276mm²)

Dies / 300mm wafer:
RV670 (55 nm) = ~292
RV770 (55 nm) = ~221 (~200 if RV770 = 276mm²)
G80 (90 nm) = 118
G92 (65 nm) = ~166
G92b (55 nm) = ~200
GT200 (65 nm) = ~89

Production cost:
GT200: ~$105
RV770: 89 / 221 * $105 = ~$42 (or $47 if RV770 = 276mm²)

I don't believe this is right. 300mm wafer is about $5K.

If the cost is 105 then 5000/105 is about 48 good dies/wafer.
If AMD has for example 80% (160) good dies (I don't know the real number) then it's 5000/160 = 31.25. That's a cost of one die.

If you go 89/221 you are applying NV yield to AMD.

 

[Wirmish]

But wafers are only one part of the total production cost...

 

[Jaaanosik]

Yes, but based on your calculation AMD has much more expensive dies than in reality.