ATI Radeon HD 4000 Series discussion

[zerazax]

I want ATI to create a one chip, highend. A two chip, single die card is not the most sustainable, nor elegent way to create a highend because its just crossfire or sli on a single card. This to me is not true singlecard performance.

Well, therein lies the big question - is R700 just CF on the board or is it different? Rumors have been suggesting its more than just CF on the board and hasn't been done before. If true, that could be huge, and hopefully thats what it is

[thenrz]

Well, therein lies the big question - is R700 just CF on the board or is it different? Rumors have been suggesting its more than just CF on the board and hasn't been done before. If true, that could be huge, and hopefully thats what it is

I'm hoping for MCM. Look at the Q6600.

[Hell Hound]

Remember people this is xtremesystems not give me value like the latest budget website like slickdeals or fatwallet.

What about Extreme Value HD4870 for 149.99us dollars.

[Shintai]

I'm hoping for MCM. Look at the Q6600.

MCM design would make no sense. You would go with a smaller combined size then with no issues, extra costs and performance penalties. CF/SLI on a board/card is due to lack of performance and heat. Just look on history with all the previous dual GPU boards when they came.

[NapalmV5]

they better do 1GB mem per gpu this time around

512MB: total joke @ high rez

i aint buying no freakn 512MB/gpu no more!!

[thenrz]

MCM design would make no sense. You would go with a smaller combined size then with no issues, extra costs and performance penalties. CF/SLI on a board/card is due to lack of performance and heat. Just look on history with all the previous dual GPU boards when they came.

OK, I see what you're saying. here's my 2c.

Dual GPU cards with a CF/SLi chip between two GPUs is no better (OK...maybe a little, but for the sake of argument) than two cards in two different PCIe slots. That needs to go. Using a dual core GPU (MCM was used loosely, all I mean is the idea of two cores in one package.) seems like the way to go...natively shared memory pools, the lack of the CF/SLi chip taking up space, much more easily designed cooling solutions, etc.

Could you expand on what you're trying to say, Shintai? I'm confused and I'd like to understand exactly what you mean.

[Shintai]

OK, I see what you're saying. here's my 2c.

Dual GPU cards with a CF/SLi chip between two GPUs is no better (OK...maybe a little, but for the sake of argument) than two cards in two different PCIe slots. That needs to go. Using a dual core GPU (MCM was used loosely, all I mean is the idea of two cores in one package.) seems like the way to go...natively shared memory pools, the lack of the CF/SLi chip taking up space, much more easily designed cooling solutions, etc.

Could you expand on what you're trying to say, Shintai? I'm confused and I'd like to understand exactly what you mean.

The design of a GPU means a MCM approach is stupid. GPUs are parallel in nature unlike CPUs. Very rough CF/SLi is the equal of a load balanced cluster for CPUs. It sucks and thats why scaling sucks usually plus latency issues.

And since GPUs dont really do well with memory mangement. A shared memory pool would be multitudes more expensive aswell.

So in short. MCM design (IF we exclude the shared memory issue)= more advanced, more transistors/die used on memory controllers, interconnection, performance scaling etc. Plus you need an interconnection that doesnt exist, not even close. Think 10-20x performance of HT3.0

In short, much more expensive nomatter what you pick.

MCM is not easier to cool. Its harder. Try cool 200-250W from a single package instead of 2.

People have a huge misunderstanding trying to think CPUs when talking/wishing about GPUs.

So this is why when performance is a dire issue. We get (G)X2 cards with 2 GPUs, 2x memory and soforth and we will in the future aswell. Nothing is changed.

And it will not improve yield either. Simply because the 2 smaller dies will have alot more critical logic that cant have defects.

This have been some runaway rumour from the start and it keep being so. People also said R680 would be MCM/dualcore whatever.

It is possible to do, but the economics makes it impossible like with so many other things.

[adamsleath]

how will rv740 compare to 3xxx series?

[w0mbat]

And how will RV730 compare to HD3k series? No one knows yet.

[AliG]

how will rv740 compare to 3xxx series?

my bet is 4670 will be at the performanc level of the 3850, or worst case 3830, a lot of the bottlenecks are improved upon to reduce/eliminate them which will unlock some of the performance of the 240 shaders, not to mention they'll have higher shader clocks than the 320 of the 3850

[trinibwoy]

I would be very surprised if the 4670 beats the 3850. It's not going to have more TMU's and it will have a shader deficit. Best chance is probably with AA enabled if there have been improvements there.

[AliG]

edit: I guess though it probably has 2 shader groups of 120 and the rv770 has 4 and so 16 and 32 work. In which case it'll still be a very good performer considering the amount of shaders, and it will still probably be at the level of the 3830

let me double check that, because if does, that'll make a big difference

[Macadamia]

I would be very surprised if the 4670 beats the 3850. It's not going to have more TMU's and it will have a shader deficit. Best chance is probably with AA enabled if there have been improvements there.

Z fillrate.

[Eastcoasthandle]

source

Default
RV770 ES clock: 625MHz
Board no.: 102-B50102-00
Revision: A12

source

Yeah, in my wet dreams

BTW: It uses 1GHz GDDR3 from Qimonda

HYB18H512321BF-10 512Mb, 16Mx32, 1.0GHz, 2.0V PG-TFBGA-136

This person states that the information is from AMD and this ES is not at full clock.

Time will tell what is true...

Also, no one seems to know about the SP count yet. It maybe either
480SPx1050MHz = 1 TFLOP
OR
800SPx625MHz = 1 TFLOP

If the 1 TFLOP is true.

In other news some are stating that "someone" over at folding forum indicated 480SP but the post was pulled from the thread. This is all rumors as far as I am concerned.

[whocaresbg]

lowered core clk may be due to huge increase in transistor density(count) .

Other interesting thing is the W1z post http://forums.techpowerup.com/showpo...9&postcount=36 .

[AliG]

umm, both posts are kinda old news, the one from folding forum is quoted from techreport and w1zzard posted that up a while ago. As for w0mbat, I wouldn't trust him if I were you, but that's just my personal choice

[rocketr2]

VR-Zone: RV770 Preliminary Performance Revealed:
http://www.vr-zone.com/articles/RV77...aled/5782.html

[thenrz]

The design of a GPU means a MCM approach is stupid. GPUs are parallel in nature unlike CPUs. Very rough CF/SLi is the equal of a load balanced cluster for CPUs. It sucks and thats why scaling sucks usually plus latency issues.

And since GPUs dont really do well with memory mangement. A shared memory pool would be multitudes more expensive aswell.

So in short. MCM design (IF we exclude the shared memory issue)= more advanced, more transistors/die used on memory controllers, interconnection, performance scaling etc. Plus you need an interconnection that doesnt exist, not even close. Think 10-20x performance of HT3.0

In short, much more expensive nomatter what you pick.

MCM is not easier to cool. Its harder. Try cool 200-250W from a single package instead of 2.

People have a huge misunderstanding trying to think CPUs when talking/wishing about GPUs.

So this is why when performance is a dire issue. We get (G)X2 cards with 2 GPUs, 2x memory and soforth and we will in the future aswell. Nothing is changed.

And it will not improve yield either. Simply because the 2 smaller dies will have alot more critical logic that cant have defects.

This have been some runaway rumour from the start and it keep being so. People also said R680 would be MCM/dualcore whatever.

It is possible to do, but the economics makes it impossible like with so many other things.

Dang man, well played. Makes perfect sense. The point about the cooling was sharp.

[LordEC911]

There are certain people on both forums you probably shouldn't listen to...

[madcho]

The blower's picture on fudzilla is a fake, AMD's logo don't have the same res with the blower.

That's not good work from them ...

[w0mbat]

The blower's picture on fudzilla is a fake, AMD's logo don't have the same res with the blower.

That's not good work from them ...

No, its not fake. Its real, but the guys @AMD changed the old ATI Radeon logo w/ the new one which came out few days ago and they did that change w/ PS, so apart from that its just a real RV770 cooler picture.

[ownage]

No, its not fake. Its real, but the guys @AMD changed the old ATI Radeon logo w/ the new one which came out few days ago and they did that change w/ PS, so apart from that its just a real RV770 cooler picture.

[ihatbs]

Olly Olly, I'm confused now.
Planning to buy a 350$ high end card before the end of july and 350$ is my max
i'm sick of buying a mid-range card every year.

anyone knows about the start prices for the new cards (both ATI & NVIDIA)

[AliG]

Then don't buy a midrange card every year, this year was a big exception with the performance outperforming the high end, and only happened due to a lack of a high end referesh until late in the game. Just get a 4870 and be happy with it or don't upgrade this year. There's not much else to do except get a consol which seems to be the way to go because of pc gaming having like no good exclusives anymore

[Nedjo]

Olly Olly, I'm confused now.
Planning to buy a 350$ high end card before the end of july and 350$ is my max
i'm sick of buying a mid-range card every year.

anyone knows about the start prices for the new cards (both ATI & NVIDIA)

in that price range, your only option will be HD4870 'cos GeForce with new GPU's will be something like 399 and 499 USD if not more!