RV770 has 800 stream processors after all...

[WorshipMe]

http://www.nordichardware.com/news,7841.html

http://forums.vr-zone.com/showthread.php?t=285701

But how did they manage to fit 800 stream processors into 230mm2 whilst still using a 55nm process.

EDIT: I still believe 480 SP's. It's almost definitely a fake slide.

[Neural Net]

http://www.nordichardware.com/news,7841.html

http://forums.vr-zone.com/showthread.php?t=285701

But how did they manage to fit 800 stream processors into 230mm2 whilst still using a 55nm process.

EDIT: I still believe 480 SP's. It's almost definitely a fake slide.

If this is true... IF... Then holy...:eek:

[chrome-187]

i dont think its fake...i think ati is awesome and are more technically advanced then nvidia atm...ati seems to be tring to innovate cards that perform as well as dont consume alot of power..nvidia on the other hand, make fast cards but the wattage is overwhelming and they are known to run hot...ati keeps things cool and keep the energy bill low with awesome performance...nvidia's you might as well get a separate electric meter for it

on the other hand Radeon HD 4870 with its 750MHz core will have 20% more processing power available and hits 1.2TFLOPS. A decent quad-core processor can do up to 100GFLOPS. This is certainly a slap in NVIDIA's face. AMD's mainstream chip can do more FLOPS than NVIDIA enthusiast chips, and AMD's coming high-end chips will pack twice the FLOPS...

[ghost_recon88]

So ATI is making a huge change in architecture, that is comparable to the change between Intel P4 and C2D. Keep bringing on the goods :D

[thideras]

i dont think its fake...i think ati is awesome and are more technically advanced then nvidia atmSo technically advanced they can defy physics? That core is extremely small. Also chrome, when you copy-paste from a site, use the quote feature ;)

I'm very interested in how these are going to do, might have to go back to ATI!

[Shiggity]

Hard to say, ATI and Nvidia usually never show their dies like AMD and Intel do.

RV700 is clearly better than GT200 at this point though, cheaper to make, runs cooler, competative performance, 55nm.

+1 ATI

[reclaimer122]

RV700 is clearly better than GT200 at this point though, cheaper to make, runs cooler, competative performance, 55nm.

But will it blend?

Hmm... now I need to make this post seem relevant...

:) Looks like a winner to me!

[chrome-187]

So technically advanced they can defy physics? That core is extremely small. Also chrome, when you copy-paste from a site, use the quote feature ;)

I'm very interested in how these are going to do, might have to go back to ATI!

lol hell yea they defy physics =D....also thideras i dont have to quote sh^t ;)

[ratbuddy]

lol hell yea they defy physics =D....also thideras i dont have to quote sh^t ;)

I don't think you're supposed to bypass the swear filter ;)

The HD2900XT had 320 SPs and the 8800GTX 'only' 128 and we know how that battle turned out. Sheer number of SPs doesn't seem to mean much..

[Neural Net]

I don't think you're supposed to bypass the swear filter ;)

The HD2900XT had 320 SPs and the 8800GTX 'only' 128 and we know how that battle turned out. Sheer number of SPs doesn't seem to mean much..

I keep reading around this forum that the SPs in ATI cards have to be divided by four in order to get the equivalent number of SPs in Nvidia cards, so 320 would equal 80 which is partly why 2900 performance was lower than the 8800GTX. I don't know if that is correct though. Anyway the little performance difference you get between the 9600GT and the 8800GT despite the 8800 having 112 SPs compared to the 9600s 64 shows us how SPs aren't the be all and end all for performance like they're made out to be.

[netmask]

Nice links. Hope those cards shake the market.

[WorshipMe]

I keep reading around this forum that the SPs in ATI cards have to be divided by four in order to get the equivalent number of SPs in Nvidia cards, so 320 would equal 80 which is partly why 2900 performance was lower than the 8800GTX.
That is BS. For a start, up until RV 770, the shader clock and core clock were linked in 1:1, meaning the shader clock would only be as high as the core clock, and ATI cards are generally faster in shader intensive games, but are held back by their lower number of TMUs and ROPs.

[chrome-187]

one thing i wonder is...why did ati not put gddr4 on the 4850 instead of the ddr3?

[ratbuddy]

one thing i wonder is...why did ati not put gddr4 on the 4850 instead of the ddr3?

My guess is performance would have been too close to the 4870 and the price would have gone up enough that noone would buy the 4850 plus they wouldn't have a card to offer for under $300..

[chrome-187]

but gddr4 is not that much more expensive then gddr3..its about the same actually =/

[MooMasster716]

but gddr4 is not that much more expensive then gddr3..its about the same actually =/
Don't worry a 3rd party manufacturer will probably have a DDR4 version available soon after launch.

[Dan0512]

Until I see the practical performance, I won't declare a winner ;-).

All this mixed up info does add to the uncertainty and excitement. Release the cards already!

And I personally doubt that the RV770 has 800 SPs, that is "R600" SPs. Like Thideras said, you can't defy the laws of physics :). If they truly did pack in so many, then they probably had a drastic change in design.

dan

[Moto7451]

Iddy bitty tiny units... and lots of them :D.

[chewietobbacca]

GDDR4 doesn't make much sense because it isn't much of an improvement over the cheaper GDDR3. GDDR5 on the other hand has 2x the effective memory rate over GDDR3 and supposedly offers actual features that may be used by R700 so it's a significant improvement over GDDR3.

R600 advertised its shaders at 320 SP's but in truth, it had 64 actual units. Each unit, however, could do 5 scalars so it was 64 * 5D = 320 SP's. Great marketing numbers, and great for folding and other GPGPU activities, but in actual games where not all 5 instructions are necessarily used, then there is no advantage. Hence synthetic benchmarks that utilize almost all of them may see huge advantages for R600, in practice it wasn't always so.

The actual increase of 64 ALU's to 160 ALU's (250% icnrease) isn't that hard to believe in a die. RV635 -> RV670 saw a huge increase in numbers and all around arcthiecture in a die increase relative to RV670 -> RV770.... and that's without needing the extra 128-bit/256-bit internal ringbus that RV670 had over the RV635.

And besides, unless we actually work for AMD and have die information, we don't know much each ALU unit actually takes up so all of us saying that its impossible based on die size don't have a clue better than anyone else.

[Kuroimaho]

And besides, unless we actually work for AMD and have die information, we don't know much each ALU unit actually takes up so all of us saying that its impossible based on die size don't have a clue better than anyone else.

Then the question would be what chew up the diespace in the previous gpus ?

[Dan0512]

http://images.bit-tech.net/news_images/2008/06/amd-hd-4850-rv770-gpu-die-shot/3.jpg

http://images.bit-tech.net/news_images/2008/06/amd-hd-4850-rv770-gpu-die-shot/rv770-1.jpg

But how did they manage to fit 800 stream processors into 230mm2 whilst still using a 55nm process.

EDIT: I still believe 480 SP's. It's almost definitely a fake slide.

The RV770 looks to be much bigger than the R670 (192 square mm), and definitely much bigger than 230 as well. To me, it looks around the same size of a G70 (7800GTX) chip, which has an area of around 340 square mm I believe:

http://www.techwarelabs.com/reviews/video/evga7800gtx512/images/pcb_s.JPG

If you take the RV670 area and add 50% to it to "simulate" 180 more SPs, then yoo would get around 300 square mm give or take. I think it's possible that it has more than 480, but if they are the same R600 shaders, then probably not up to 800.

dan

[ratbuddy]

Just a heads up, I think that card in the first picture needs more thermal paste :)

[WorshipMe]

Just a heads up, I think that card in the first picture needs more thermal paste :)

Yep, and they also forgot to squirt some nice conductive stuff over the PCB.

[vixro]

What a sloppy thermal paste removal job! Looks like they used their fingers.

[WorshipMe]

I have to admit that die is a decent size, this might just be true.

[MooMasster716]

Is it just me or does that RV770 card look rather bare to any of you.

[Dan0512]

Is it just me or does that RV770 card look rather bare to any of you.

It's a simple 256 bit bus design that doesn't need a complex PCB compared to the HD2900 XT and the competition.

The chip also consumes "little" power hence the simple power circuitry.

dan

[Shell]

Are those 1.1GHz Qmonda GDDR3 chips on that 4850?

If so... WHY NOT GDDR4? It's not any more expensive and would not starve the thing as much.

[ratbuddy]

Are those 1.1GHz Qmonda GDDR3 chips on that 4850?

If so... WHY NOT GDDR4? It's not any more expensive and would not starve the thing as much.

The core itself probably isn't fast enough to need any faster memory.

[Shell]

The core itself probably isn't fast enough to need any faster memory. I can wholeheartedly disagree with that.

If GDDR3 is retarding the 8800GT/GTS(G92) even with excellent memory management, how can a significantly faster GPU have memory that's only a little faster without consequence?

Isn't ATi is claiming 1.3x performance with GDDR5?
Sounds significant, and I bet you could clock the core much higher without the memory being a significant... retarding factor.

[ratbuddy]

I can wholeheartedly disagree with that.

If GDDR3 is retarding the 8800GT/GTS(G92) even with excellent memory management, how can a significantly faster GPU have memory that's only a little faster without consequence?

Isn't ATi is claiming 1.3x performance with GDDR5?
Sounds significant, and I bet you could clock the core much higher without the memory being a significant... retarding factor.

4870 is supposed to be 1.3x faster than 9800GTX.. For a card that's rumored to be on par with 8800GTS (4850), GDDR3 ain't gonna hold it back..

[juane414]

My guess is that the 4850 will be marketed as middle-end and the 4870 will be marketed as high-end. If the 4850 had GDDR4 or GDDR5 then there wouldn't be much of a reason to buy the 4870.

[Dan0512]

My guess is that the 4850 will be marketed as middle-end and the 4870 will be marketed as high-end. If the 4850 had GDDR4 or GDDR5 then there wouldn't be much of a reason to buy the 4870.

My thoughts exactly. Ati is giving manufactures a lot of freedom regarding these cards, so I think we'll be seeing DDR4 versions in the market as well.

dan

[dealmaster]

So is it confirmed that the 4870 will only have 1 6-pin PCI-E power connector?

[MR-FIX-IT]

I'm just waiting to see who comes out with a dual core or quad core GPU...............first