R600 nice read

[Nelsieus]

Originally posted by: ronnn
Are you in this thread to discuss the r600 or bash ati and pump nvidia? If you are saying the inquirer is frequently wrong - what was your first clue? If you think the r600 sucks, just say why as a history lesson in the 7800gtx 512mb is about something I don't remember.

I'm not sure how you can take "I'm not saying R600 isn't going to be truly amazing, because I really have that faith in ATI" to be bashing ATI and proclaiming R600 as sucking.

Nelsieus

 

[Nelsieus]

Originally posted by: redbox

I agree on most parts except this. The 7800gtx 512 came 6 months after the first G70. To expect them to have enough yeild in their chips especially one as complex as G80 in 4 months is a little optomistic. I don't doubt they will be producing one, I just don't see them producing one in that short of a time table.

As far as the 7950GX2, they where only able to produce that card because of G71's lower power consumption along with lower heat out put. True we could see another die shrink for G80, but not even close to the time table of Feb. The 7950GX2 was a year after the original G70's. The only thing I can see Nvidia putting together in time for R600 is an 8800gt i.e. another cut down of the gtx.

redbox

True. Although I'm keeping in mind the fact that G80 has supposedly been finished and working for some time now. As far as Geforce 7950GX2, I agree it would nearly be impossible for nVidia to bolt two 8800GTS together. Perhaps, though, they won't need that much power to combat R600, and that it will only be a matter of increasing clocks a few hundred MHz. But then again, R600 might blow us all away, just like G80 did, and nVidia might have a more serious problem on their hands.

Originally posted by: Cookier Monster96vec4 shaders sound impressive. R600 might probably have the edge in theorectical shader performance if that was true. But in a real performance sense, G80s scalar shaders is utlised much more than a vec4 shader. Although vec4s do much more work per cycle, we are talking about 100% utlisation of the scalar shaders in G80. You will need R600 to be clocked around 700mhz to match this. Then even so, the R600 shader design could be vec4+scalar found in the Xenos.

96Vec4 would be impressive, but not likely, imo. 64vec4 sounds more in line with other rumors. In which case, I think it just may come down to clock speeds. 128 scaler @ 1350 ~ 64vec4 @675 only with more effiency and higher utilisation. If R600 is 64vec4, I think it would need to be clocked around 700-750MHz, like you said, to be competitive.

But there's other factors to take in than my simply little theory, so I'm not making any predictions quite yet.

Nelsieus

 

[Munky]

I think it's more than 64 shaders. Not only do those shaders have to share geometry and fragment shading, but 64 would not be doubling of current gen specs, and IMO not enough to entail a 512-bit bus. The reason I'm expecting a 512-bit bus is because they're not likely to change the memory subsystem for a refresh, and so it has to be a forward-looking design.

 

[Cookie Monster]

Originally posted by: munky
I think it's more than 64 shaders. Not only do those shaders have to share geometry and fragment shading, but 64 would not be doubling of current gen specs, and IMO not enough to entail a 512-bit bus. The reason I'm expecting a 512-bit bus is because they're not likely to change the memory subsystem for a refresh, and so it has to be a forward-looking design.

Do you think its a good idea to use 512bit memory interface? would it be worth it? Is that much bandwidth needed right now?

I guess those are some of the key questions regarding the 512bit rumour. I do also believe they will use the memory controller found in the R5x0. 16 Memory chips and the 512bit sounds like a nightmare for PCB design though.

 

[josh6079]

Do you think its a good idea to use 512bit memory interface? would it be worth it? Is that much bandwidth needed right now?

Perhaps not "right now" but "right now" there isn't an R600. Who knows if it will need that amount of bandwidth to carry out its features. Heck, who even knows what features it will have? I've only heard the hardware aspects of the card, not the feature set ones. What kind of AA is it going to push? SSAA? Higher levels of MSAA? What's its AF going to be like? Quantum Physics? etc.

 

[Munky]

Originally posted by: Cookie Monster

Originally posted by: munky
I think it's more than 64 shaders. Not only do those shaders have to share geometry and fragment shading, but 64 would not be doubling of current gen specs, and IMO not enough to entail a 512-bit bus. The reason I'm expecting a 512-bit bus is because they're not likely to change the memory subsystem for a refresh, and so it has to be a forward-looking design.

Do you think its a good idea to use 512bit memory interface? would it be worth it? Is that much bandwidth needed right now?

I guess those are some of the key questions regarding the 512bit rumour. I do also believe they will use the memory controller found in the R5x0. 16 Memory chips and the 512bit sounds like a nightmare for PCB design though.

IMO, they could get away with a 256-bit bus right now if they use fast ddr4 mem and have more shader power than the g80. But eventually it will be a disadvantage when the g80 switches to ddr4 mem, so I would think they'll use a 512-bit bus now to have much more bandwidth available for future derivatives of the r600.

 

[Cookie Monster]

But wouldnt that cripple the R600 initially? Didnt nVIDIA chose to use GDDR3 compared to GDDR4 as it is widely available without having to pay a price premium nor face the prospect of having memory shortages?

 

[josh6079]

But wouldnt that cripple the R600 initially? Didnt nVIDIA chose to use GDDR3 compared to GDDR4 as it is widely available without having to pay a price premium nor face the prospect of having memory shortages?

It's got to become widely available somehow. I'd imagine with DAMiT initiating the use of it and placing the demand for it, there will eventually be a better supply of GDDR4 chips. Besides, they used it with their X1950XTX, I don't see why they wouldn't with their new flagship.

Whether it cripples the R600 initially, who knows. Maybe DAMiT is aware of a greater supply of GDDR4 than we are. If not, I would expect DAMiT to have lower yeilds initially but all cards do. The G80 right now probably has lower yields compared to what it will be yielding once its price falls a little.

 

[Cookie Monster]

The GDDR4 used by the X1950XTX is the slowest of the bunch. Munky was referring to high speed GDDR4 memory.

This situation sort of reminds me of the nv30 days. It would be horrible if the memory used by R600 hurt it in terms of quantity, similiar to the 5800 ultra and its 1000mhz GDDR2 chips.

 

[redbox]

Originally posted by: Cookie Monster
The GDDR4 used by the X1950XTX is the slowest of the bunch. Munky was referring to high speed GDDR4 memory.

This situation sort of reminds me of the nv30 days. It would be horrible if the memory used by R600 hurt it in terms of quantity, similiar to the 5800 ultra and its 1000mhz GDDR2 chips.

IMO ATI would be better suited using the 512 bit bus slaping in 1gb or even the "slow" GDDR4 stuff. Then pushing for the highest GDDR4. This does a number of things for them. One it makes sure they don't run into problems like Nvidia does with low mem numbers i.e. 5800 Ultra and 7800gtx 512. 2 it still gives some energy savings over GDDR3 which this generation will most likely need. 3 it gives some room for further expanision. When it comes down to it really the core on the R600 is going to be the deciding factor.

 

[Munky]

Originally posted by: Cookie Monster
The GDDR4 used by the X1950XTX is the slowest of the bunch. Munky was referring to high speed GDDR4 memory.

This situation sort of reminds me of the nv30 days. It would be horrible if the memory used by R600 hurt it in terms of quantity, similiar to the 5800 ultra and its 1000mhz GDDR2 chips.

That's another reason why a 512-bit bus would be a better option, in that it would not depend on high speed ddr4 mem for good performance.

 

[Cookie Monster]

Originally posted by: munky

Originally posted by: Cookie Monster
The GDDR4 used by the X1950XTX is the slowest of the bunch. Munky was referring to high speed GDDR4 memory.

This situation sort of reminds me of the nv30 days. It would be horrible if the memory used by R600 hurt it in terms of quantity, similiar to the 5800 ultra and its 1000mhz GDDR2 chips.

That's another reason why a 512-bit bus would be a better option, in that it would not depend on high speed ddr4 mem for good performance.

Then again, they could be going with the slower GDDR4 knowing that with the increased bus would still have more bandwidth.

So, what do you guys think will come new with R600? I know that it wont be as big of a jump as R4x0-->R5x0 or G7x-->G80, but what will ATi bring new to the table? Obviously AF will be improved. New AA modes, and hopefully some SSAA. AVIVO and memory controller will probably stay the same. The composition engine could be included in the R600 die as well.

Can anyone make a prediction on the die size if R600 is a unified shader archtiecture with 96 vec4 shaders including other extras? (They are using 80nm or 65nm? , im thinking 80nm)

 

[josh6079]

I know that it wont be as big of a jump as R4x0-->R5x0 or G7x-->G80...

I don't know about that. If it will be faster than the G80, how can its R580-->R600 not be as big of a jump from G71-->G80?

 

[Cookie Monster]

Originally posted by: josh6079

I know that it wont be as big of a jump as R4x0-->R5x0 or G7x-->G80...

I don't know about that. If it will be faster than the G80, how can its R580-->R600 not be as big of a jump from G71-->G80?

Im talking about it in a architectural sense. Not in performance sense. Just like how the X1800XT was only 5% faster in performance with 7800GTX (nv47, which indeed is very old). Just because its fast doesnt mean its a big jump. The G80 from what i can see is MUCH more radical compared to a G7x. The architecture itself is very intriguing, and the decisions nVIDIA made to this architecture is very smart. e.g sticking to scalar shader instead of vec4 which gave nVIDIA to do many things including the ability to easily clock their shaders beyond 1 ghz. (not to mention utlisation is at 100%)

When it comes to architecture, GPU companies cant afford to refresh their architecture from ground up in a 12month cycle. (These are mostly in the works though) It just doesnt make sense to do so because you are throwing away the millions of R&D spent on the previous architecture. You could rather be more conservative and develop that previous architecture into a more efficent form while slowly working on the next big jump.

In this case, ATi has both spent R&D on R500 and R5x0. The R600 architectural layout was probably finalised some time ago. To me, it makes sense to take the best of both worlds in this case. Definately ATi will use the memory controller, including their ring bus technology (and the ultra thread dispatcher) while utlising their experience to make a unified shader architecture GPU gained from the R500/Xenos project.

Example : The 8500 was the foundation for the R300, which later on grew to its full potential in the shape of the X850XTPE which out performed the NV4x which in architectural sense was more advanced.

To me, the R5x0 has barely showed its full potential. Compared to the G7x, its very much more advanced.

Im thinking the R700 will be the real next gen from ATi. R600 could most probably be the final architecture using elements of R5x0 and R500 before jumping into the "true" DX 10.1 card therefore R700.

But this is all specualation. I could be wrong. If this is the derivative of the R400, then we could expect something entirely different compared with current gen ATi GPUs.

note - some people in b3d are saying that the A1 revision of the R600 failed to work. Its also said to be bigger than the G80 even using 80nm, and it might weigh just over 750 million transistors. Take it with a pinch of salt.

Thats posting for one day.

 

[Munky]

My prediction is that it will be quite a big jump from the r5xx cards, similar to the g80 (in terms of performance and design). I'm guessing 600-700M transistors, on 80nm process, with 96 vec4 shaders and a 512-bit bus. And I'd also like to have SSAA, although I'm not so sure it will happen.

 

[hardwareking]

Actually i heard that the problem with revison A1 was that 4xAA or greater wouldn't work.
They won't/can't reduce the die size.All they can do is trim down the PCB and employ better cooling.

 

[Cookie Monster]

I was wondering, is there any site who took the heat spreader off and took a die shot of G80? Would be great for comparison.

 

[Nightmare225]

 

[Warren21]

For all you speculators out there.. I'm gonna be a little more conservative.

I'm gonna give credit to most of what the INQ is saying here with these specs:

Expect the final product to be close to:

700-800 MHz (Maybe 750) Core clock (unified architecture, so the shaders do not have independant clock like G80)
16 ROPS (according to the Inq, remember G80 only has 24)
64 Vec4 shaders (equal to 256 scalar shaders remember, so 64 vec4's is no push over)
512-bit external/1024-bit internal ringbus
1 GB GDDR4 from 1.0-1.25 GHz (2.0-2.5GHz effective, this is on the highest end model.. I think XTX/XT = GDDR4, XT/XL=GDDR3)
~550M transistors on 80nm
~225W Power draw tops (PCIe x16 BUs + 2 x 6-pin PEG connectors = 75W x 3)
Inq said PCB = X1950 XTX size, so 9" around in PCB length

 

[Cookie Monster]

Is 700~850mhz core even possible? a unified shader architecture based on vec4 is much more complex than scalar based. Im also hearing that the R600 will use vec4+scalar, just like the Xenos.

 

[Gstanfor]

Well, given the noises emanating from AMD about GP-GPU and stream processors I'd say that the decision (if true) to use a vector rather than scalar based design is more than a little puzzling, given that scalar is better suited for GP-GPU (it can be more completely utilized than vector per clock cycle and is easier to compile for).

 

[redbox]

Originally posted by: Gstanfor
Well, given the noises emanating from AMD about GP-GPU and stream processors I'd say that the decision (if true) to use a vector rather than scalar based design is more than a little puzzling, given that scalar is better suited for GP-GPU (it can be more completely utilized than vector per clock cycle and is easier to compile for).

What if it used both vector and scalar? Would that be a better fit for the GP-GPU?

 

[Munky]

Originally posted by: redbox

Originally posted by: Gstanfor
Well, given the noises emanating from AMD about GP-GPU and stream processors I'd say that the decision (if true) to use a vector rather than scalar based design is more than a little puzzling, given that scalar is better suited for GP-GPU (it can be more completely utilized than vector per clock cycle and is easier to compile for).

What if it used both vector and scalar? Would that be a better fit for the GP-GPU?

It's better, but not as good as pure scalar shaders. On the other hand, the scalar ALU's require more complex thread scheduling hardware to keep the ALU's busy. Also, dynamic branching performance has a big impact on gpgpu performance, and that is more dependent on the sceduling logic than the actual ALU's themselves.

 

[Cookie Monster]

Leaked R600 die shots

R520 and R600 compared

It looks like R600 isnt too far off from its launch as leaks are spreading. However, notice that the R600 is the same size in terms of the package. Therefore many b3d members think that there wont be enough pins for 512bit memory bus interface and its going to be around 534 million transistors. (note that the R600 is 80nm).

This leads me to believe that R600 is going to be 256bit using faster GDDR4 memory. You have to keep in mind that GDDR4 at 1.4ghz (256bit) offers more bandwidth than what the G80 offers even with its 384bit. Theres also an possibility of the R600 using 384bit.

 

[Kromis]

Originally posted by: Cookie Monster
Leaked R600 die shots

R520 and R600 compared

It looks like R600 isnt too far off from its launch as leaks are spreading. However, notice that the R600 is the same size in terms of the package. Therefore many b3d members think that there wont be enough pins for 512bit memory bus interface and its going to be around 534 million transistors. (note that the R600 is 80nm).

This leads me to believe that R600 is going to be 256bit using faster GDDR4 memory. You have to keep in mind that GDDR4 at 1.4ghz (256bit) offers more bandwidth than what the G80 offers even with its 384bit. Theres also an possibility of the R600 using 384bit.

Nice!