The LAST R600 Rumours & Speculation Thread

[CarstenS]

Well, I'm laughing. That is totally, totally nonsensical.

Sorry, must have been the language barrier.
This sentence "Der R600 ist einfach ein verdoppelter R580 mit den nötigen D3D10-Erweiterungen?" translates into "The R600 simply is a doubled R580 with the necessary D3D10-extensions". Extensions obviously not meaning anything OpenGL-like but additions to the hardware to make it D3D10-compatible. (Don't worry - it is marked as speculation though)

Obviously i am not level with the B3D-amount of competence, but as far as i've been told, the pixel shader ALUs in R580 come pretty close to be "almost D3D10-capable" in the same sense that Geforce FX was pretty close to SM3.

 

[Rebel44]

Anyway we should soon be able to find some reliable test which will confirm/deny level505 tests. I hope that X2800XTX will have hardlaunch because I want to buy new rig in april and I hope that pricis could be around 600 euro.

 

[AlexV]

Obviously i am not level with the B3D-amount of competence, but as far as i've been told, the pixel shader ALUs in R580 come pretty close to be "almost D3D10-capable" in the same sense that Geforce FX was pretty close to SM3.

Mehh, quite close means nothing if you`re quite close to jumping over a huge rift, but don`t quite make the jump, eh?I think that the evolution is more from Xenos, rather then R580, which would be quite logical-Xenos is indeed very close to being a DX10 equivalent part.

 

[Geo]

Anyway we should soon be able to find some reliable test which will confirm/deny level505 tests.

Only if the universe has a sense of humour. :smile: The drivers will have been 4 months less mature. It won't be production silicon. And I personally doubt that if, by some miracle, they had a real board, that it'd be a fully functional one (compared to what is sold to the public) for reasons that wouldn't cause it to fail WHQL (so it'd have been fine for driver testing). So any result from level505 that bears a passing resemblance to what is released in March will be 1) Proof of the existance of a time machine or 2) More proof that God has an evil sense of humour.

 

[seahawk]

So far I gather this together:

64 4-Way SIMD Unified Shaders “64 pixel unified processorâ€￾, 128 Shader Operations/Cycle, 32 TMUs, 16 ROPs, 512 bit Memory Controller, GDDR4 at 1.1 GHz clock speed, total bandwidth 140 GB/s on GDDR4, total video memory 1024 MB, 814MHz GPU, 32FP internal processing, over 700 million transistors, RingBus 1024bit internal, A15 silicon.

nVidia can process only one operation per clock cycle while R600 can do four.
Also nVidia 24 ROP operating and frequency of 575 MHz provides 13.4 billion:
ATI at 16 ROP operating and frequency of 814 MHz can provide very close to nVidia:

I mean what exactly we are talking here, 4 lanes of Freeway traffic driving 50MPH vs. 1 lane of Freeway traffic driving at 200MPH. “Just a questionâ€￾

To fill up your specs a bit.

64 4+1 Unified shaders for a total of 128 Shader Ops per clock.

Comparing the processing power to the G80 however is difficult. Sure in the best case 64 Vec4+1 ALUs would outperfrom 8 VEC16 (aka 128 SPs) However if R600 is fed with scalar ops only, then it can only do 2, now consider the higher clock of G80.

Now if you have 2D Vec to process (not uncommon for texturing) G80 can do 2, R600 one.

 

[Robbitop]

but as far as i've been told, the pixel shader ALUs in R580 come pretty close to be "almost D3D10-capable" in the same sense that Geforce FX was pretty close to SM3.

I doubt that. AFAIK they don't support:
-Int format
-IEE-754 compliance (specific approximation rules, specific calculation accuracy, specific configuration of the FP format which differs from the former configuration)
-fMAD with FP32
-inf/nan with spec. approx
-fMAD FP16 w/out denorm
-support of non-number-construction terms (my engl suxx I know)
ectpp

Why should they support these transistor consuming things for D3D9, if they were not neccessary?

 

[Robbitop]

Now if you have 2D Vec to process (not uncommon for texturing) G80 can do 2, R600 one.

Per Vec4/5 ALU? I don't see any reason why ATI should not support a 2 : x split, which would allow to do 2x Vec2 operations as well per clock and Vec4/5 (whatever this will be) ALU.
But to aquire 2x differenct Vec2 operations for the same quad won't be happen so often.

 

[CarstenS]

I doubt that. AFAIK they don't support:
-Int format
-IEE-754 compliance (specific approximation rules, specific calculation accuracy, specific configuration of the FP format which differs from the former configuration)
-fMAD with FP32
-inf/nan with spec. approx
-fMAD FP16 w/out denorm
-support of non-number-construction terms (my engl suxx I know)
ectpp

Why should they support these transistor consuming things for D3D9, if they were not neccessary?

What's your guess then how these features do weigh in terms of transistor count? We're speculating here about 720M transistors von R600 - pretty close to double the number of R580 (782M). Now subtract eight Vertexshaders, and "one half" the necessary 2D-part - you'd end up pretty close to 720M IMO.

And while you/we are at guessing: What's your guess why R580 needed about 42% higher t-count than G71 despite having fewer individual units? The massive jump nvidia had to take relates IMO to a great percentage to decoupling their TMUs, improving AF and adopting the (almost) D3D10-mandatory threading concept including a massive registerfile - which Ati already did.

 

[DegustatoR]

Considering how NV shocked everyone with G80

I wouldn't say that NV shocked _everyone_

 

[Hanners]

Sorry, must have been the language barrier.
This sentence "Der R600 ist einfach ein verdoppelter R580 mit den nötigen D3D10-Erweiterungen?" translates into "The R600 simply is a doubled R580 with the necessary D3D10-extensions". Extensions obviously not meaning anything OpenGL-like but additions to the hardware to make it D3D10-compatible. (Don't worry - it is marked as speculation though)

That doesn't fit in with the fact we know R600 will be unified, for starters...

 

[Ailuros]

I wouldn't say that NV shocked _everyone_

You aren't exactly the average reader either

 

[Robbitop]

What's your guess then how these features do weigh in terms of transistor count? We're speculating here about 720M transistors von R600 - pretty close to double the number of R580 (782M). Now subtract eight Vertexshaders, and "one half" the necessary 2D-part - you'd end up pretty close to 720M IMO.

But we don't know the count of transistors, are we?
I think this is a too much superficial point of view. Just adding transistors without the knowledge the details of the cost about the realisation of D3D10 specs is kind of senseless. Anyway I can't provide this kind of knowledge because I'm no chip architect.

G80 shows that ALUs for D3D10 can be very expensive. But the reason for that is complex.

And while you/we are at guessing: What's your guess why R580 needed about 42% higher t-count than G71 despite having fewer individual units? The massive jump nvidia had to take relates IMO to a great percentage to decoupling their TMUs, improving AF and adopting the (almost) D3D10-mandatory threading concept including a massive registerfile - which Ati already did.

That's architecture related. The Transistor count is comprised of many points. One would be the cache design, one would be the ringbus, one would be the threading concept and so on. As you said R5xx is pretty advanced in it's kind of aranging the workflow to the particular units. Much control logic is needed. NV4x is pretty static. Of cource we only know a abrasive view of the logical construction of these archichitecture but without knowing about the real conctruction behind.
R580 has more ALUs than G71. This would be expensive too.

I guess G80 is not a decoupled advanced NV4x. They had to design much logic from the start.

In the end I personally guess, R600 is more related to R400/C1 than it is to R5xx. And I don't know how much any detail would cost.

 

[Geo]

To fill up your specs a bit.

64 4+1 Unified shaders for a total of 128 Shader Ops per clock.

Oh, is that what they meant by "128 Shader Ops per clock"?

 

[XMAN26]

he posted those benchmarks because he now believes that its true - he has to have some good reason to say that because in january he posted that its fake and none of us would change opinion without reason.

The fact that none of the numbers have changed or that the Oblivion numbers are way off, doesn't seem to throw flags up for you? I think you are more guilible than he is. Some one is trying hard to force feed these bogus numbers and hard.

 

[CarstenS]

That doesn't fit in with the fact we know R600 will be unified, for starters...

I was merely talking about doubling Pixel-Shaders. 48*2=96, a number Dave Orton did mention a while ago as a possibility...

 

[CarstenS]

I guess G80 is not a decoupled advanced NV4x. They had to design much logic from the start.

In the end I personally guess, R600 is more related to R400/C1 than it is to R5xx. And I don't know how much any detail would cost.

I agree with you on the first one - the example was, how the total of the D3D10 feature set did cost. The number of ALUs did not even go up in total, considering eight VS- and 24(*2) PS-ALUs. Granted, they're not comparable in any sane way.

But i think with R580/R600 it's quite different. Many of Nvidias additional transistors were used up simply by the fact that they had to switch towards a more threaded approach, which Ati already had done - hence the large number of transistors on R5xx-products.

 

[Arty]

Digitimes

Upcoming launch of AMD-ATI R600 helps ease IC substrate oversupply, says paper

The upcoming dual-branded AMD-ATI graphics chip, the R600, which is made on 80nm, is helping ease the oversupply of IC substrates, according to a Chinese-language Commercial Times report, which noted that the chip is expected to be officially announced on either March 8 or 9.

The paper said that orders for IC substrates started to be released in late January and the ramp up in demand was evident in the January sales record of major substrate makers, including Nanya Printed Circuit Board (NPC), Phoenix Precision Technology (PPT) and Kinsus Interconnect Technology.

 

[neliz]

sdbear.com has the latest on the drive inf files.

http://www.sdbear.com/2/lib/200702/07/20070207259.htm

This one also talks about RV610 and 630 supporting a maximum of 4 devices.
http://www.sdbear.com/2/lib/200702/08/20070208182.htm

It's obvious someone in china has hold of the drivers.

 

[turtle]

Wouldn't it be sad if the drivers that include those strings (probably cats that will ship with the card) are WHQL before nvidia gets their equal out the door? Granted, they probably won't show up in regular Cats until 7.3/7.4, but looking at nvidia ATM, that still might beat them.

Here's shouting a loud prayer to a silent god the R6xx series works with AFR/OGL/D3D10/CF right out the gate on Vista.

 

[KimB]

Wouldn't it be sad if the drivers that include those strings (probably cats that will ship with the card) are WHQL before nvidia gets their equal out the door?

They would need to get the hardware out the door first.