*OFFICIAL* R700 thread

JPB

Diamond Member
Jul 4, 2005
4,064
89
91
Fudzilla Link

That is why R680 makes sense


It looks like ATI wants to go multi core and this time it can put more smaller cores on a single chip. You can expect that the high-end R700 might end up with more than four smaller cores that will shape up this card.

This is why R680, Radeon 3870 X2 is extremely important for ATI?s future, as it looks like that the future is really multicore. We still don?t know if G100 uses the same approach, but we would not be surprised.

R700XT, the high-end version of the chip will use four or more smaller cores to reach the performance crown and, in this case, you need an excellent Crossfire, multichip driver.

R680 is actually a warm-up product, as ATI will test the Crossfire X with two, three and four cores, and this will give them predictions about the future. R700 is still scheduled for middle of 2008.

______________

Radeon 3870X2 uses Alternate Frame Rendering

If you wondered how CrossfireX works, now we have an answer for you. We asked few AMD product managers and they confirmed that 3870 X2 dual RV670 card or Crossfire with four 38x0 cards works with the help of Alternate Frames Rendering magic.

Alternate frame rendering is quite simple model, where the first chip renders the first frame, the second chip renders the second one and so on.

If you have as many as four cards, each card will render two images, something like card one renders the first and forth while card two renders the second and fifth frame, the third card renders the third and sixth frame and so on.

This should give AMD better performance and leaves space for eight way crossfire. In 3Dmark06 debugging mode we had a chance to see that four cards are rendering eight images virtually at the same time.

_______________

R700 is a 45 nanometre chip

AMD and ATI to go for it


Nvidia just went for 65 nanometre, ATI is at 55 nanometre and it looks like the R700 might be the world?s first 45 nanometre graphics chip.

A 45 nanometre chip is at least 30 percent smaller than 55nm and its more than 50 percent smaller than 65 nanometre chip and at the same time this means that it will be way cooler than any 65 or 55 nanometre chip.

This also means lower heat dissipation and higher clock speeds, but we still don?t know G100 can be the same deal and it can be made of many smaller cores as well. This is how it is possible to put just a bit less than 300 million transistors on a 72 sq mm die.

________________

R700 mini core is 72 sq mm

R700 looks like an interesting chip as it will change a lot of things. Nvidia currently talk about clusters but most of these chips are blocks of functional units divided in these so called clusters.

ATI will do a step forward and will introduce its mini core that will be one small part of a larger chip. ATI can use just one R700 core to get an entry level chip, two cores will make a mainstream card and four R700 cores should take care of the high end.

A single R700 core has just below 300 million transistors and you can do the math that even four R700 cores means over a billion transistors and a lot of graphics processing power.

_________________

High end R700XT has four+ mini cores

Four+ times R700

From what we´ve seen, the R700 is a small single core and if you want a high end you will have to put more of these together.

With lower dissipation and 45 nanometer it will be possible to have four, or even six, smaller cores. The R700 single core will be similar to what we seen in the guise of the RV670, but it is still too early to have too many details.

The new chip will support DirectX 10.1 and Shader model 4.1, among other things. Imagine a chip that can have three or even more times performance of RV670, and this might give you a glimpse of what is coming.

__________________

R700 to carry Radeon HD 4000 name

A whole new thing

There is a big chance that the R700 generation will carry the Radeon HD 4000 generation name, but there is also a chance that ATI will lose the HD prefix or completely change the name.

Since the 3870 is already out, the new R700 marchitecture at 45 nm would deserve a much higher number, and that is likely to happen. It is still too early for a final conclusion, as ATI likes to play around with the names; but at least there is a big change.

AMD internally toyed with that name, but they did the same with Radeon 2950 as a possible name for RV670. Radeon HD 4x00 sounds quite nice, doesn't it?

___________________

R700XT to touch 2 Teraflops

4X theoretical performance of RV670

The next generation of ATI´s latest greatest multi core chip has a chance to be close to or to even break the 2 teraflop mark.

The RV670XT is close to 500 gigaflops, while R700XT might even get as fast as 1.9 teraflops, if not even more. Depending on the final clock the new card might even do 2 Teraflops, and this will certainly speed up games or GPGPU calculations, up to four times.

Imagine if you put two of these in Crossfire, you might even get to 4 Teraflops; but now we are really pushing it.

____________________

R700 is a low, mid and high-end chip

Mini cores

There won't be any RV chips anymore, as the entire R7x0 generation is now based on the R700 mini core. It is quite simple, a single core on a single die is low-end, two cores on a single die is mid-range, while four or more cores on a single die is high-end.

ATI will have to put these R700 cores together and we still don't know how they are going to make the chips "talk" to each other; but it will use the same architecture chip but multiplied in the mid and high-end market.

Sounds interesting doesn't it ? It also explains why ATI is pushing so hard for a Quad CrossfireX driver.

_____________________

Tech Report Article

Modern graphics processors rely on extensive amounts of parallelism to get calculations done as quickly as possible, but those GPUs are still monolithic chips designed with a given number of stream processors, texture units, render back-ends, and the like, each depending on their performance grade. However, the folks at Fudzilla say they have it on good authority that AMD will challenge that paradigm with its next-generation graphics processor, code-named R700. According to Fudzilla, low-end, mid-range, and high-end R700 cards will all have GPUs with varying numbers of identical R700 cores. To determine speed grades, AMD will simply outfit higher-end cards with more R700 cores.

AMD's top-of-the-line R700 product, for instance, will supposedly have four or more R700 cores in one die and will be able to crunch nearly two trillion floating point operations per second, or teraFLOPS. By contrast, Fudzilla explains that the existing Radeon HD 3870 is in the 500 gigaFLOPS range. FLOPS don't tell the whole story, of course, but high-end R700 cards will be an order of magnitude faster than existing products if those numbers are even remotely accurate.

Interestingly, this rumor sounds similar to information that recently seeped out about Intel's Larrabee project. Larrabee is expected to be a discrete, game-worthy Intel graphics processor scheduled for the not-too-distant future, and an Intel presentation nabbed by Beyond3D in April suggests Larrabee products will be based on multiple, small "throughput cores." A diagram showed a chip based on ten of those cores with a shared pool of 4MB of cache.

 

thilanliyan

Lifer
Jun 21, 2005
12,040
2,256
126
I'm so tempted to go Crossfire...since it's so cheap now...just for the hell of it. Dang it I need a new motherboard too. But then I think...the only game that really requires it is Crysis...and I don't really wanna play that anyway. COD4, Gears of War, and Timeshift run very well at 1920x1080. Sighh...what to do...what to do?? :)

What I'm wondering is...will people that already have Crossfire not have much of an upgrade when the new cards come out?? (ie.) Will R700 be faster than say 4x3850/3870 Crossfire?? If that IS the case then why not just go Crossfire or SLI right now I suppose.
 

Munky

Diamond Member
Feb 5, 2005
9,372
0
76
I really hope by "multi-core" they mean something more elegant than slapping 4 gpu's onto the board, each with its own memory space, and having them depend on Crossfire to make it all work.
 

BadRobot

Senior member
May 25, 2007
547
0
0
@thilan the high performace card yet to be released SHOULD have more vram...if it doesn't they need slapped...
 

chizow

Diamond Member
Jun 26, 2001
9,537
2
0
Originally posted by: munky
I really hope by "multi-core" they mean something more elegant than slapping 4 gpu's onto the board, each with its own memory space, and having them depend on Crossfire to make it all work.

Sadly I think this will be the greatest limiting factor for any multi-GPU set-up. There's really only so much VRAM you can slap onto a PCB. Unless they use denser modules, which would drive up prices exponentially.

But as it is now, I'm pretty sure total VRAM is split amongst # of GPUs, so if you have 1GB and 2 GPU than you'd effectively have 512MB per core. With 4 cores, that'd be 2GB on a single PCB.....

What I'm not sure about is how Alternate Frame Rendering handles this compared to a single GPU. For example, it seems 4 cores would render 8 frames simultaneously, with each frame needing its own frame buffer. I think 1 core only stores 3 frames simultaneously (hence Triple Buffering), which means 1 core in a multi-GPU would probably need a little less RAM than a single GPU.
 

Munky

Diamond Member
Feb 5, 2005
9,372
0
76
Why would 4 cores need to render 8 frames? The way I see it, each core render 1 frame, so you'd get 4 frames simultaneously. Gpu's use multiple buffers, but the actual rendering only get's done in the front buffer, and then the data is just copied to the back buffer. Unless there's something about multi-gpu configurations that I'm not aware of...
 

nismotigerwvu

Golden Member
May 13, 2004
1,568
33
91
Perhaps having access to AMD's engineers will help in creating shared resources. The first thing that comes to mind is perhaps a large pool of texure/general purpose memory (which with AFR no more than 1 chip at a time should be trying access) and then individual pools for frame buffers for each core (much akin to the old voodoo days, only with shared texture space to increase efficiency). Have some logic chip that keeps things sorted out and transperent to the other chips. I know this is a very (VERY VERY VERY) over simplified view, but this method does seem feasible.
 

chizow

Diamond Member
Jun 26, 2001
9,537
2
0
Originally posted by: munky
Why would 4 cores need to render 8 frames? The way I see it, each core render 1 frame, so you'd get 4 frames simultaneously. Gpu's use multiple buffers, but the actual rendering only get's done in the front buffer, and then the data is just copied to the back buffer. Unless there's something about multi-gpu configurations that I'm not aware of...

The article actually stated 8 frames being rendered by 4 cores...... I'd guess they're rendering 8 frames simultaneously to maximize utilization/efficiency.

I'm not saying a multi-core GPU would have problems with rendering power, as it should have that in spades. The point of my post was that shared VRAM on a single PCB would be the issue to keep all 4 of those cores happy.

The reason I think its going to be an issue is the same reason we see with current gen cards where lower VRAM variants perform significantly worst than versions with more RAM (320/640MB GTS) and even 512MB GT vs 640MB+ GTS/GTX at higher resolutions or with AA enabled. If any individual core ran out of VRAM for whatever # of frames it was expected to store in VRAM you'd get performance issues similar to what you see with a single GPU that runs out of VRAM.
 

Munky

Diamond Member
Feb 5, 2005
9,372
0
76
That's what I have in mind as the ideal multi-core gpu. Having multiple "worker" cores, each with its own cluster of shaders and texture units, but all running off a central memory controller, so they can share the available memory space and not rely purely on software to arbitrate the work between the cores.
 

Munky

Diamond Member
Feb 5, 2005
9,372
0
76
Originally posted by: chizow
Originally posted by: munky
Why would 4 cores need to render 8 frames? The way I see it, each core render 1 frame, so you'd get 4 frames simultaneously. Gpu's use multiple buffers, but the actual rendering only get's done in the front buffer, and then the data is just copied to the back buffer. Unless there's something about multi-gpu configurations that I'm not aware of...

The article actually stated 8 frames being rendered by 4 cores...... I'd guess they're rendering 8 frames simultaneously to maximize utilization/efficiency.

I'm not saying a multi-core GPU would have problems with rendering power, as it should have that in spades. The point of my post was that shared VRAM on a single PCB would be the issue to keep all 4 of those cores happy.

The reason I think its going to be an issue is the same reason we see with current gen cards where lower VRAM variants perform significantly worst than versions with more RAM (320/640MB GTS) and even 512MB GT vs 640MB+ GTS/GTX at higher resolutions or with AA enabled. If any individual core ran out of VRAM for whatever # of frames it was expected to store in VRAM you'd get performance issues similar to what you see with a single GPU that runs out of VRAM.

That's true, the current multi-gpu implementations are not making the best use of total available memory.
 

TanisHalfElven

Diamond Member
Jun 29, 2001
3,512
0
76
hey maybe the "cores" aren't complete. there just set of shaders each connected to a main chip that does the load balancing.

eg.
1. controll chip
2. texture unit chips
3. shader chips
4. other part i know nothing about

1+2+3+4= whole gpu.
maybe its a away to increase yeilds by making smaller chips in a very large quantities.

then a low end card could get less shader and texture units than the high ends one.

 

Sylvanas

Diamond Member
Jan 20, 2004
3,752
0
0
Originally posted by: tanishalfelven
hey maybe the "cores" aren't complete. there just set of shaders each connected to a main chip that does the load balancing.

eg.
1. controll chip
2. texture unit chips
3. shader chips
4. other part i know nothing about

1+2+3+4= whole gpu.
maybe its a away to increase yeilds by making smaller chips in a very large quantities.

then a low end card could get less shader and texture units than the high ends one.

Isn't that kinda how the PS3's cell processor is structured? In that it doesn't have 8 individual 'cores' so to speak- just different parts of the silicon dedicated to a certain task (correct me if I'm wrong). Whatever the case it seems AMD/ATI and Nvidia are coming to the conclusion Intel and AMD came to a couple of years ago, multicore is the future.
 

Munky

Diamond Member
Feb 5, 2005
9,372
0
76
Originally posted by: Sylvanas
Originally posted by: tanishalfelven
hey maybe the "cores" aren't complete. there just set of shaders each connected to a main chip that does the load balancing.

eg.
1. controll chip
2. texture unit chips
3. shader chips
4. other part i know nothing about

1+2+3+4= whole gpu.
maybe its a away to increase yeilds by making smaller chips in a very large quantities.

then a low end card could get less shader and texture units than the high ends one.

Isn't that kinda how the PS3's cell processor is structured? In that it doesn't have 8 individual 'cores' so to speak- just different parts of the silicon dedicated to a certain task (correct me if I'm wrong). Whatever the case it seems AMD/ATI and Nvidia are coming to the conclusion Intel and AMD came to a couple of years ago, multicore is the future.

Depends on what is considered a gpu "core." Gpu's are already massively parallel by design, with a gpu being divided into quads or "clusters", and each quad contains an array of "pipes" or shaders working in parallel. The use of multiple gpu "cores" may just be for lower defect rates in production, and lower cost of manufacturing.
 

tuteja1986

Diamond Member
Jun 1, 2005
3,676
0
0
A's Quantum theory on GPU !
1+1+1+1 = 4 + 1+1+1+1 = Anti gup meet normal gpu = both explode = create some much performancer that it runs Crysis 2560x1600 @ Ultra high setting 64xAA / 64xAF :! 10000000FPS :!

 

Jeff7181

Lifer
Aug 21, 2002
18,368
11
81
Originally posted by: munky
Originally posted by: Sylvanas
Originally posted by: tanishalfelven
hey maybe the "cores" aren't complete. there just set of shaders each connected to a main chip that does the load balancing.

eg.
1. controll chip
2. texture unit chips
3. shader chips
4. other part i know nothing about

1+2+3+4= whole gpu.
maybe its a away to increase yeilds by making smaller chips in a very large quantities.

then a low end card could get less shader and texture units than the high ends one.

Isn't that kinda how the PS3's cell processor is structured? In that it doesn't have 8 individual 'cores' so to speak- just different parts of the silicon dedicated to a certain task (correct me if I'm wrong). Whatever the case it seems AMD/ATI and Nvidia are coming to the conclusion Intel and AMD came to a couple of years ago, multicore is the future.

Depends on what is considered a gpu "core." Gpu's are already massively parallel by design, with a gpu being divided into quads or "clusters", and each quad contains an array of "pipes" or shaders working in parallel. The use of multiple gpu "cores" may just be for lower defect rates in production, and lower cost of manufacturing.

That's what I thought right away. A GPU is already massively parallel. A multi-core GPU would exists for two reasons. One... to lower manufacturing cost... high end GPU's with 4 cores, low end GPU's with 1 core. Or two... for marketing purpsoes. "ZOMG I bought a multi-core GPU to go with my multi-core CPU. Now I need multi-core RAM and a multi-core mouse so I can pwn everyone at TF2!"
 

BadRobot

Senior member
May 25, 2007
547
0
0
Originally posted by: Jeff7181
Originally posted by: munky
Originally posted by: Sylvanas
Originally posted by: tanishalfelven
hey maybe the "cores" aren't complete. there just set of shaders each connected to a main chip that does the load balancing.

eg.
1. controll chip
2. texture unit chips
3. shader chips
4. other part i know nothing about

1+2+3+4= whole gpu.
maybe its a away to increase yeilds by making smaller chips in a very large quantities.

then a low end card could get less shader and texture units than the high ends one.

Isn't that kinda how the PS3's cell processor is structured? In that it doesn't have 8 individual 'cores' so to speak- just different parts of the silicon dedicated to a certain task (correct me if I'm wrong). Whatever the case it seems AMD/ATI and Nvidia are coming to the conclusion Intel and AMD came to a couple of years ago, multicore is the future.

Depends on what is considered a gpu "core." Gpu's are already massively parallel by design, with a gpu being divided into quads or "clusters", and each quad contains an array of "pipes" or shaders working in parallel. The use of multiple gpu "cores" may just be for lower defect rates in production, and lower cost of manufacturing.

That's what I thought right away. A GPU is already massively parallel. A multi-core GPU would exists for two reasons. One... to lower manufacturing cost... high end GPU's with 4 cores, low end GPU's with 1 core. Or two... for marketing purpsoes. "ZOMG I bought a multi-core GPU to go with my multi-core CPU. Now I need multi-core RAM and a multi-core mouse so I can pwn everyone at TF2!"

Hmm where do i buy this multi-core mouse you speak of? I could always use more pwnage in tf2!
 

biostud

Lifer
Feb 27, 2003
19,746
6,828
136
shared vram and memory subsytem seems like one of the major inefficiencies of current SLI/crossfire
 

TanisHalfElven

Diamond Member
Jun 29, 2001
3,512
0
76
Originally posted by: BadRobot
Originally posted by: Jeff7181
Originally posted by: munky
Originally posted by: Sylvanas
Originally posted by: tanishalfelven
hey maybe the "cores" aren't complete. there just set of shaders each connected to a main chip that does the load balancing.

eg.
1. controll chip
2. texture unit chips
3. shader chips
4. other part i know nothing about

1+2+3+4= whole gpu.
maybe its a away to increase yeilds by making smaller chips in a very large quantities.

then a low end card could get less shader and texture units than the high ends one.

Isn't that kinda how the PS3's cell processor is structured? In that it doesn't have 8 individual 'cores' so to speak- just different parts of the silicon dedicated to a certain task (correct me if I'm wrong). Whatever the case it seems AMD/ATI and Nvidia are coming to the conclusion Intel and AMD came to a couple of years ago, multicore is the future.

Depends on what is considered a gpu "core." Gpu's are already massively parallel by design, with a gpu being divided into quads or "clusters", and each quad contains an array of "pipes" or shaders working in parallel. The use of multiple gpu "cores" may just be for lower defect rates in production, and lower cost of manufacturing.

That's what I thought right away. A GPU is already massively parallel. A multi-core GPU would exists for two reasons. One... to lower manufacturing cost... high end GPU's with 4 cores, low end GPU's with 1 core. Or two... for marketing purpsoes. "ZOMG I bought a multi-core GPU to go with my multi-core CPU. Now I need multi-core RAM and a multi-core mouse so I can pwn everyone at TF2!"

Hmm where do i buy this multi-core mouse you speak of? I could always use more pwnage in tf2!

noegg.com
 

dreddfunk

Senior member
Jun 30, 2005
358
0
0
JPB - thanks, those were interesting reads. This type of modular design was very much where I thought the industry would go eventually, though I had no idea when it would become practical. There are just so many price-points to fill that having a core that can scale from 1 to 4+ on a die would save a lot on R&D and manufacturing.

Would anyone that knows the engineering better care to explain some of the problems they would have to solve (memory allotment, access and core coordination, etc.) in order to make this work?
 

firewolfsm

Golden Member
Oct 16, 2005
1,848
29
91
It's not practical for AMD to make this multicore card anything like the 3870X2, that card is for a very small market and the price will reflect that.

I see this more or less like Intel's current quad cores. Nothing will really seem different to the end user, but the manufacturer can cut down on production costs and increase yields. The chips might even be set under one IHS, or, like AMD's Fusion, sharing some components for efficiency. This also means that the rest of the PCB wouldn't be affected (ie, one set of memory core the card, normal drivers)

Which brings up the issue. It IS less efficient to produce a card like this, there will be extra logic needed to support each individual core, as opposed to running more shaders under one core, and one cache. This means more transistors and possibly more heat. The cost of the transistors will be offset by cutting down on variety, which is the advantage of this approach. Simplifying the process by only manufacturing one product will make the chips cheaper for AMD.
 

taltamir

Lifer
Mar 21, 2004
13,576
6
76
Originally posted by: tanishalfelven
hey maybe the "cores" aren't complete. there just set of shaders each connected to a main chip that does the load balancing.

eg.
1. controll chip
2. texture unit chips
3. shader chips
4. other part i know nothing about

1+2+3+4= whole gpu.
maybe its a away to increase yeilds by making smaller chips in a very large quantities.

then a low end card could get less shader and texture units than the high ends one.

interesting. Yesturday I was just thinking how AMD should have designed the phenom this way.


@sylvanas, no it is very different then a cell processor, in that a cell processor is a single die containing 8 cores. While this idea is splitting one core into parts and then pairing up parts as needed.


I was also thinking about taking it in the GPU market as well...
There is a good chance that me and you aren't the only ones thinking about it, and that this is what the AMD engineers are making with the RV700. This sounds like it has some potential.
 

ConstipatedVigilante

Diamond Member
Feb 22, 2006
7,670
1
0
Thanks for the updates, OP. This is really quite interesting. Everyone has been saying that multiple chips on a card would be super impractical due to high heat, memory, etc., but I've heard few people mention a multi-core chip.