VRZone: R680 to hit 20K in 3dMark06

[zerazax]

It even has a leaked slide! Looks like the dual RV670 configuration is indeed true. If a single card is 1.5x faster than 2 2900XT's in CF then the teraflop beast age is indeed here.

http://www.vr-zone.com/articles/R680_To_Score_20K_In_3DMark06/5325.html

Now if Intel doesn't get SLI, Nvidia cards in Vista still have driver issues, and the R680 actually fixes the AA and other performance issues, getting one of these and being able to CF it w/ Intel boards might be worth a shot.

[RPGWiZaRD]

For AMD's sake I hope they'd reach 20k+ 3Dmark06 because that's a logical "must-reach" goal, comparing how 8800GTX/Ultra does now I don't think it will be any prob for next gen flagship NVIDIA card to reach that either if NVIDIA's playing the game right. :D At least DAAMIT have to put the bar a lot higher than they did with R600.

We even heard faintly that R680 could be AMD's ambitious plan to integrate two RV670 into a single die, if not on the same package.

Hmm I wish but seeing how AMD loves this solution if we compare CPU market I wouldn't yet deny it won't happen. But I wonder how they'd keep that chip cool under such circumstances. :D Guess it would mean return to the "vacuum cleaners", 4000rpm 80 - 92mm fans or sth lol.

NVIDIA's reign has been long enough now, I hope all the best for DAAMIT this time around and prays it gets the performance crown this time around so we can get a little more balanced GFX market. Same goes for CPU market unfortunately.

[Lightman]

That looks scary!
ATI might have a good chance to beat nVidia this round! That would return balance to GFX market for another 1-2 years at least!:)

PS. Anyone is wondering what would be score with Quad-Crossfire using these boards?? :shocked:

[b0bd0le]

i hope this is true

because weren't there rumors that the successor to the 8800GTX would have double the processing power?

[zanzabar]

ati just took a year off from their 8-10years of supremacy over nvidia, the rage pro -1900xtx ati had a better card befor or withing 2 months of every nvidia relies and for most of teh time atis last gen midrange beat atis top cards (the fx5k and 9x00 days oh how i loved those days when a grafix card was an afterthought to a cpu and ram)

[Yakyb]

any idea on release date?

[artissco]

sick powaaa :up: :shocked:

[Eastcoasthandle]

It seems that the information posted a few months ago might be true. The poster was right about the RV670 (months before released) and now looks like he was right about the dual RV670. Now I like to see if it will be 2 gpus put together like the C2Ds.

[Solarfall]

i will believe when i see it :rolleyes: i mean this is just speculation now, there are no benchmarks published. yet

[Shintai]

Lets just hope this aint a hype like the R600 :(

[Kasparz]

Lets hope it have software voltage adjust and it won't suck your power bill.

[perkam]

That looks scary!
ATI might have a good chance to beat nVidia this round! Nvidia is quickly becoming the Intel of the GPU industry...it's just too hard to sneak up on it nowadays.

Perkam

[Californian]

I hope this won't need 1200W PSU.
Ther poweconsumption is turning into something ridiculous year after year.
And that cooling solution is better to be efficient & quiet. That's something ATI hasn't sone in years but hope lives on :)

[zanzabar]

It seems that the information posted a few months ago might be true. The poster was right about the RV670 (months before released) and now looks like he was right about the dual RV670. Now I like to see if it will be 2 gpus put together like the C2Ds.


wouldent that be like the pentium D with 2 dies on 1 chip, the core2 moved to the point were amd was where it had 2 cores on 1 die

[zerazax]

Well if the RV670's cooler design and the 55nm shrink really does result in much lower power consumption and heat, then a dual RV670 core card might require less power and put out less heat than even a single 2900XT...

Anyways, I'm just glad that finally there is some semblance of competition between ATI and NVidia again.

2950Pro is making Nvidia release an 8800GT
R680 might be the counter to the next Nvidia high end (G90/92/whatever)

And anyways, I'm sure ATI doesn't mind having this stuff leaked out right now because even if Nvidia beats ATI to the punch and gets a new high end out before Christmas, the mere news that ATI might have a competitive if not better card out within a few months can draw a few customers away. Heck, it worked for the R600 but at least this time, the goals are a lot loftier.

And yeah, that user who posted the info on the RV670 months ago and the R680 as well has been pretty spot on so far... RV670 as a new upper mid-range as well as R680 being the high-end as a dual RV670 core card.

[zerazax]

wouldent that be like the pentium D with 2 dies on 1 chip, the core2 moved to the point were amd was where it had 2 cores on 1 die

No one knows for sure. It could be 2 separate dies / cores on a single board for all we know or a 7950GX2 style configuration.

R700 is supposedly like Core 2 Duo with 2 cores on 1 die...

[RPGWiZaRD]

Well if the RV670's cooler design and the 55nm shrink really does result in much lower power consumption and heat, then a dual RV670 core card might require less power and put out less heat than even a single 2900XT...

I have a hard time believing that, 55nm would improve things a lot but it won't be enough for a dual gpu solution to draw less than a single X2900XT that's for sure but perhaps 230~250W is realistic but how to keep such a beast cool in acceptable ways (price, noise, temp factors taken in mind) for the global market is beyond my imagination unless ATI has done something we're not aware about yet.

Well time will tell as usual.

[zerazax]

Well power consumption with 2 cores will probably be close to the 2900XT maybe a tad higher

Heat wise? If its two cores on one board, a dual slot cooler seems to be in order. If it is like the 7950GX2 with basically 2 RV670 single slots stuck together, that might work too.

Either way, this looks like it will be the first card to truly use PCIE 2.0's increased power slots + 8 pin PCIE connectors to the fullest... and Quad-Fire too if you use 2 of these cards in CF

[OBR]

RV670 looks like very good chip ... finally! And this 2xRV670 monster i hope will be in my rig ...

[Eastcoasthandle]

You have to remember most of the heat/power is created/needed for the gpu and PCB not the GPU alone. With only 1 PCB the amount of heat/power created/needed should be within reason, making it more effcient. My only concern is the gpu clock rate (will using 2 gpus make the video card more efficient). With 2 gpus (IE C2D) a high clock shouldn't be needed IMO. For example for a dual core gpu solution at a 3D clock rate of 600/2400 providing 20,000 Marks in 3D06 would be impressive versus 900/2400 providing 20,000 Marks in 3D06.

[wedfighter]

can't wait to see and try it :)

[RPGWiZaRD]

You have to remember most of the heat/power is created/needed for the gpu and PCB not the GPU alone. With only 1 PCB the amount of heat/power created/needed should be within reason, making it more effcient. My only concern is the gpu clock rate (will using 2 gpus make the video card more efficient). With 2 gpus (IE C2D) a high clock shouldn't be needed IMO. For example for a dual core gpu solution at a 3D clock rate of 600/2400 providing 20,000 Marks in 3D06 would be impressive versus 900/2400 providing 20,000 Marks in 3D06.

Good point there.

I wonder how exactly the 2 GPUs would communicate/work together if anything at all with each other, there can be huge variance in power/heat output here depending how the architecture is designed.

[BeardyMan]

This could be possible yes, if they also managed to bump up cpu score with some beta drivers it's very very possible.

[Brian MP5T]

SUBscribble

[flopper]

first time I like what ati has to offer since x1950xtx

if its true, it be a great card to have.
finnaly one to upgrade to.

[highoctane]

Lets hope there's more to it than a 3dmark score, even the current cards are good 3dmark performers.

[MAS]

PS. Anyone is wondering what would be score with Quad-Crossfire using these boards?? :shocked:

you need at least 4.5GHz quad-penryn (unfortunately 65nm Phenom X4 can't reach such freq.) and pci-e 2.0 to see the mighty of 2xRV680 ;)

[quadnad]

Any word on release?

[STaRGaZeR]

first time I like what ati has to offer since x1950xtx

if its true, it be a great card to have.
finnaly one to upgrade to.

Me think the same. x1950XTX was sooooa wonderful card... 2900XT is faster in some situations, but the lack of AA perfomance sucks badly. It´s even slower than 1950 sometimes, go figure.

[Revv23]

i need to replace my x1900.... it need to come out before xmas or i might have to jump to nvidia.

[Origin_Unknown]

You have to remember most of the heat/power is created/needed for the gpu and PCB not the GPU alone. With only 1 PCB the amount of heat/power created/needed should be within reason, making it more effcient. My only concern is the gpu clock rate (will using 2 gpus make the video card more efficient). With 2 gpus (IE C2D) a high clock shouldn't be needed IMO. For example for a dual core gpu solution at a 3D clock rate of 600/2400 providing 20,000 Marks in 3D06 would be impressive versus 900/2400 providing 20,000 Marks in 3D06.

the link says the core is going to be 800+ mhz

[Morais]

For AMD's sake I hope they'd reach 20k+ 3Dmark06 because that's a logical "must-reach" goal, comparing how 8800GTX/Ultra does now I don't think it will be any prob for next gen flagship NVIDIA card to reach that either if NVIDIA's playing the game right. :D At least DAAMIT have to put the bar a lot higher than they did with R600.



Hmm I wish but seeing how AMD loves this solution if we compare CPU market I wouldn't yet deny it won't happen. But I wonder how they'd keep that chip cool under such circumstances. :D Guess it would mean return to the "vacuum cleaners", 4000rpm 80 - 92mm fans or sth lol.

NVIDIA's reign has been long enough now, I hope all the best for DAAMIT this time around and prays it gets the performance crown this time around so we can get a little more balanced GFX market. Same goes for CPU market unfortunately.

I agree completely

[Shintai]

wouldent that be like the pentium D with 2 dies on 1 chip, the core2 moved to the point were amd was where it had 2 cores on 1 die

GPUs dont work like CPUs. CPUs can share memory. GPUs cant.

So you will need double the memory for 2 cores. Or some seriously super I/O that would kill the cost.

[GoThr3k]

GPUs dont work like CPUs. CPUs can share memory. GPUs cant.

So you will need double the memory for 2 cores. Or some seriously super I/O that would kill the cost.

why can't gpu's share memory?

[Shintai]

why can't gpu's share memory?

Because they aint designed for it in their nature.

The best way to think of it might be as each GPU as a multicore/SMP system, and a remote GPU as another box. Cluster if you like, or CF/SLI in modern terms.

[Origin_Unknown]

Because they aint designed for it in their nature.

The best way to think of it might be as each GPU as a multicore/SMP system, and a remote GPU as another box. Cluster if you like, or CF/SLI in modern terms.

cf / sli still doesnt share memory iirc does it? it just allocates one gpu / memory to one half of the screen and the other pair to the other half of the screen?

[cadaveca]

Ringbus nodes allow for shared framebuffers. One ringbus node is connected to PCI-E interface, another is used for Crossfire, so don't go discounting shared framebuffer just yet. In fact, ebcause of the way the image is composted, and the lack of an external composting engine, BOTH sli and Crossfire have a bit of shared buffer, nV uses I/O chip for that tho, which, of course, could then send data to either card via the SLI bridges...making for a slow write, for sure, but Crossfire does not use a "middleman" for communication, so each write could potentially be direct to frambuffer of other card.

Tri-SLI connector spans both bridges of 8800's...

[Flinch]

After the hype surrounding the 2900 launch, I'm extremely skeptical. At this point, I have to see it to believe it.

[Zytek_Fan]

Is the resurgence of ATi coming? :D

Working RV670, Phenom system pixellated (http://www.theinquirer.net/gb/inquirer/news/2007/10/06/working-rv670-phenom-system)
:D


edit: If anyone complains about the article mentioning Fudzilla, remember that VR-Zone is a highly credible source.

[theteamaqua]

sounds good, if its the best card at launch, i hope its not over like $600 .... so Q1 08?? sigh .. still a long time to go ...

[Sneil]

Is the resurgence of ATi coming?


man i sure hope so for their sake. I'd hate to see them left in the dust once again, look what it's done, we haven't seen a refresh/next gen card from both sides in over a year! remember when it used to be every six months or so?

[Origin_Unknown]

man i sure hope so for their sake. I'd hate to see them left in the dust once again, look what it's done, we haven't seen a refresh/next gen card from both sides in over a year! remember when it used to be every six months or so?

I'm glad its not every 6 months.

Much like the xbox and ps3 programers should learn to use what they've got and program games accordingly. Likewise, the gfx makers should use the extra time to make the cards better instead of rushing them out every 6 months

[RPGWiZaRD]

Man those INQ articles sure hypes you up about RV670 but yea either 2950 Pro or 8800GT will be my next card for sure that's decided but now it remains to see which. But if those things about power consumption, performance figures etc. are true this will be a hell of a midrange card.

Is AMD finally going a similiar approach as NVIDIA and try compete with efficient cards? Usually ATI cards are way more power hungry but looks like RV670 might change the concept of ATI (pressure from AMD folks?) and that would be a very welcomed addon to me. Judging by the cooler designs on both 8800GT and 2950 Pro it looks like both will produce about as little heat, man I can't wait to see how these babies will clock. Can't stop drooling over 55/65nm, a bit above 100W power consumption and ~$250 price tag. I was like thinking I prolly end up paying 350 EUR or so for my next card that would be 70% or faster than my current overclocked card but now it seems more like 200 EUR will be enough, that's music to my ears for a poor student like me addicted to hardware. xD

Man it will be an exciting Q4, 8800GT vs 2950 Pro battle alone I think is more exciting than Penryn and Phenom launch! :p:

[Jimmer411]

Lets hope it have software voltage adjust and it won't suck your power bill.

And how much did installing a 2900XT raise your powerbill?

[Shintai]

cf / sli still doesnt share memory iirc does it? it just allocates one gpu / memory to one half of the screen and the other pair to the other half of the screen?

Precisely. Thats also why I compared it to a cluster. Load balancing if you like.

Ringbus nodes allow for shared framebuffers. One ringbus node is connected to PCI-E interface, another is used for Crossfire, so don't go discounting shared framebuffer just yet. In fact, ebcause of the way the image is composted, and the lack of an external composting engine, BOTH sli and Crossfire have a bit of shared buffer, nV uses I/O chip for that tho, which, of course, could then send data to either card via the SLI bridges...making for a slow write, for sure, but Crossfire does not use a "middleman" for communication, so each write could potentially be direct to frambuffer of other card.

Tri-SLI connector spans both bridges of 8800's...

I/O for CF and PCIe got nothing to do with memory. They still cant share memory. Same reason we also see the 2600XTx2 cards with "double" memory.

SLi and CF dont share anything as such. Their buffers are just indentical loaded with the same textures etc.

So in short, no.

And there is no Tri-SLI today. Only Quad and Dual. And the quad with...quad memory.

[BeardyMan]

quad sli? dual sli?lol:ROTF:

sli = 2 no? :rolleyes:

[Shintai]

quad sli? dual sli?lol:ROTF:

sli = 2 no? :rolleyes:

If SLI as a definition was 2, then quad SLI would be 8 GPUs, yet its just 4. :rolleyes:

SLI and Crossfire is a method, not a number. Its just used to refer to 2 by default.

[BeardyMan]

If SLI as a definition was 2, then quad SLI would be 8 GPUs, yet its just 4. :rolleyes:

SLI and Crossfire is a method, not a number. Its just used to refer to 2 by default.

i smell contradiction :ROTF:

[jdam]

any reason why they are moving back to GDDR3 in those specs? i mean the x1950xtx allready used GDDR4.

[Shintai]

any reason why they are moving back to GDDR3 in those specs? i mean the x1950xtx allready used GDDR4.

GDDR3 is cheaper and within the speedrange it seems. It also offers better thermals. GDDR4 is a semi flop besides the speed. And most put their effort on GDDR5 while skipping GDDR4.

[cadaveca]

Precisely. Thats also why I compared it to a cluster. Load balancing if you like.



I/O for CF and PCIe got nothing to do with memory. They still cant share memory. Same reason we also see the 2600XTx2 cards with "double" memory.

SLi and CF dont share anything as such. Their buffers are just indentical loaded with the same textures etc.

So in short, no.

And there is no Tri-SLI today. Only Quad and Dual. And the quad with...quad memory.

LoL, well, I've seen Tri-Sli, and yes, not in market, but that doesn't mean it doesn't exist.:fact:

Secondly, due to the organization if the ringbus, it is MORE than possible, as the final frame is merged on the "master" card's framebuffer...how did it get there? it jsut magically appeared? is that why Vista Crossfire sucks?:lol2:

Ringbus = front/back buffer access from same controller. Data from one bank can be accessed and sent to another, thereby allowing fully rendered frames to be post processed...in fact, Bioshock does this, for a fact, as the lighting is done by drawing two frames and tehn interpolating the difference for the final frame. As such, and because of lighting, the only way this can work is if the framebuffers can access each other.

As I stated earlier, nV uses I/O chip to manage this for SLi, ATI doesn't, so ATI CAN share framebuffer, nV cannot with the same speed, due to the extra interconnects.

Tri-SLi uses a "dual" bridge that uses both connecters on the 8800's...this bridge is one single unit, unlike ATi's dual-ribbon solution.

[Shintai]

LoL, well, I've seen Tri-Sli, and yes, not in market, but that doesn't mean it doesn't exist.:fact:

Secondly, due to the organization if the ringbus, it is MORE than possible, as the final frame is merged on the "master" card's framebuffer...how did it get there? it jsut magically appeared? is that why Vista Crossfire sucks?:lol2:

Ringbus = front/back buffer access from same controller. Data from one bank can be accessed and sent to another, thereby allowing fully rendered frames to be post processed...in fact, Bioshock does this, for a fact, as the lighting is done by drawing two frames and tehn interpolating the difference for the final frame. As such, and because of lighting, the only way this can work is if the framebuffers can access each other.

As I stated earlier, nV uses I/O chip to manage this for SLi, ATI doesn't, so ATI CAN share framebuffer, nV cannot with the same speed, due to the extra interconnects.

Tri-SLi uses a "dual" bridge that uses both connecters on the 8800's...this bridge is one single unit, unlike ATi's dual-ribbon solution.

Thety do not share memory at all. One card simply get the result from whatever mode they render in from the other and direct passes it to the output.

Or maybe those 2 12bit channels (bridge) can do more? No..

Its another wish dream. Just like 2 RV670 dies on a pcs of silicon in the original FUD part. It makes no sense at all due to a GPUs super parallel design. It would be a disadvantage rather than advantage.

[tsahi]

you need at least 4.5GHz quad-penryn (unfortunately 65nm Phenom X4 can't reach such freq.) and pci-e 2.0 to see the mighty of 2xRV680 ;)

Why my dear fellow, that is the reason we have 4*4, two socket F mb:

http://www.asus.com/products.aspx?l1=3&l2=82&l3=489&l4=0&model=1530&modelmenu=1

You see, all hope is not lost yet.

[cadaveca]

Or maybe those 2 12bit channels (bridge) can do more? No..

Its another wish dream. Just like 2 RV670 dies on a pcs of silicon in the original FUD part. It makes no sense at all due to a GPUs super parallel design. It would be a disadvantage rather than advantage.

12-bit @ what speed? and what about pci-e interconnect?:lol2:


Um, for Rv760 dual... if they have full 512bit ringbus controller on each gpu, front buffer/back buffer seperated, each gpu could access any point of the ringbus of the other gpu. NUMA, anyone? And, becuase of the gpu's highly parralell nature, it MORE than makes sense, as when one part goes idle, seeing how it can do any operation that any other part can do, it can do extra work passed off from one of the other "pipes". You could think of it akin to Core2's shared L2...because the actual functions would be VERY similar, yet more massively parrallel.


This ringbus would be the perfect inter-gpu link...proprietary, and definately has the nessecary bandwidth. INterestingly enough, R580 has this same kind of functionality, which allowed it to do AA+HDR...already processed frames are reprocessed. It also had 512-bit framebuffer with half turned off.:fact: They have more experince in practical usage of such a memory control scheme than most image, IMHO.

[tsahi]

We still have no games that can utilize all 4 cores. Single and dual threaded performance is the key for the performance of all available games and that won't change in near future. You might have a zillion quadcores on a board, but they won't perform faster than a single quadcore.

You mentioned a quad core, we will also have a tri core, in any case a quad
or tri 670 set up with a good cpu (dual, tri or quad or 2 * quad) will be a nice set up.

Whem those products will come out we will see where the bottle neck is
with new gaming engines.

[zerazax]

Well the good news is that it looks like Nvidia rushing out the 8800GT is because ATI's 2950 was pressuring them to do so...

Wow, quite some time since Nvidia has had to rush anything out because of ATI

[Shintai]

12-bit @ what speed? and what about pci-e interconnect?:lol2:


Um, for Rv760 dual... if they have full 512bit ringbus controller on each gpu, front buffer/back buffer seperated, each gpu could access any point of the ringbus of the other gpu. NUMA, anyone? And, becuase of the gpu's highly parralell nature, it MORE than makes sense, as when one part goes idle, seeing how it can do any operation that any other part can do, it can do extra work passed off from one of the other "pipes". You could think of it akin to Core2's shared L2...because the actual functions would be VERY similar, yet more massively parrallel.


This ringbus would be the perfect inter-gpu link...proprietary, and definately has the nessecary bandwidth. INterestingly enough, R580 has this same kind of functionality, which allowed it to do AA+HDR...already processed frames are reprocessed. It also had 512-bit framebuffer with half turned off.:fact: They have more experince in practical usage of such a memory control scheme than most image, IMHO.


Do you just guess or?

The PCIe aint used for ANYTHING between the 2 cards when a bridge is used. The only thing the PCIe is used for is the change of textures and such. There is no intercommunication as such.

One card with half its memory bandwidth turned off is how again compared to sharing memory between 2 GPUs? Thats right, nothing.

The 2 12bit channels are in GPU terms slow, but used to transfer the final frame or half frame etc. In other words are are talking about something with the need for a few Gbits at max.

NUMA? Please...you gotta be joking. Do you even know what NUMA gives in terms of performance when you need the memory on the other CPU? Thats right, one huge penalty. For a GPU something like HT3.0 would be useless at best. You would need something 20x or more faster than HT3.0 for a highend card.

Its parallel nature is also what limits it. All the benefits are already used and the extreme bandwidth needed is not practical. You would need a higher interbandwidth than to ones memory. We talk 120GB/sec+. The ringbus is INTERNAL and cant handle the external I/O.

And do NOT compare a CPU with a GPU. A Core 2 Duo would at best be a 2 pipeline GPU. The same parallel reason is why multicore or 2 cores on a die is USELESS for a GPU.

Its all nice if we sit and imagine the best wishes again. But then we just gonna end up with another K10/R600 hype.

[Shintai]

Well the good news is that it looks like Nvidia rushing out the 8800GT is because ATI's 2950 was pressuring them to do so...

Wow, quite some time since Nvidia has had to rush anything out because of ATI

I disagree. nVidia changed to a 1 year cycle. And looking back on G80. Its time for a new card in november.

[Zytek_Fan]

I just say don't get your hopes up because you can be disappointed...

[cadaveca]

Its all nice if we sit and imagine the best wishes again. But then we just gonna end up with another K10/R600 hype.
If you were right, say, and pci-e was of no use, then there would be no increase in Crossfire scores from 16x16 VS 8x8, however, even X1950 gets almost 1000 points in '06 from such...why? because of the intercommunication.


Crossfire does not communicate over a set bus, it is designed to work over ANY bus that links the two cards. AS such, additional links can be used for additional work, and this is seen in a 16x16 config. You can also think a bit here...there's no way that if communication was entirely by the bridges that X1950 Software Crossfire would work...but it does...oh, and by the way, only in a 16x16 config, almost like the card's were designed for such...:lol2::stick::lol2:

In regards to comparing a cpu and a gpu, I'm not as far off as you think, however your own perspective on the gpu is too coarse...each pipeline within a gpu is akin to a cpu, and R600 has 64 of them. the use of Core2 was merely an example...L2 data can be accessed by either core, and cores communicate over the FSB.

You could compare the frambuffer as the L2, and the pci-e peer-to-peer link as the FSB(or SPI link, or Bridge link), but of course, because it doesn't suit your purposes, you won't. I know better tho...

NUMA is Non-Unified Memory Architechture, and yes, current implementations are slow, however they are not ringbusses. HT 3.0? bandwidth?:lol2: 20.xGB?:lol2: How's 50GB, in 256-bit ringbus...oh, that's right, more than twice the bandwidth.:stick: And enough, if as you say they merely pass rendered frame data back and forth, but then, in the instance of SFR and HDR...what happens?

Gotta come up with better than that...nothing you have mentioned really prevents dual-gpu configs from sharing memory access, nor does it prevent dual-gpu configs on a single substrate. They are obstacles, sure, but they are obstacles already overcome.:fact:


However, I must admit, that maybe none of these features get used...as you say, AMD isn't known right now for been wholey upfront about thier products

[SparkyJJO]

I just say don't get your hopes up because you can be disappointed...

I'm not because I don't care all that much. I can not afford a single upgrade anyway so what does all this matter to me? :p: :(

[Zytek_Fan]

I'm not because I don't care all that much. I can not afford a single upgrade anyway so what does all this matter to me? :p: :(

I'm excited for R680, but I don't want to get my hopes up too much. I want an all AMD system that rocks :(

[Shintai]

If you were right, say, and pci-e was of no use, then there would be no increase in Crossfire scores from 16x16 VS 8x8, however, even X1950 gets almost 1000 points in '06 from such...why? because of the intercommunication.


Crossfire does not communicate over a set bus, it is designed to work over ANY bus that links the two cards. AS such, additional links can be used for additional work, and this is seen in a 16x16 config. You can also think a bit here...there's no way that if communication was entirely by the bridges that X1950 Software Crossfire would work...but it does...oh, and by the way, only in a 16x16 config, almost like the card's were designed for such...:lol2::stick::lol2:

In regards to comparing a cpu and a gpu, I'm not as far off as you think, however your own perspective on the gpu is too coarse...each pipeline within a cpu is akin to a cpu, and R600 has 64 of them. the use of Core2 was merely an example...L2 data can be accessed by either core, and cores communicate over the FSB.

You could compare the frambuffer as the L2, and the pci-e peer-to-peer link as the FSB(or SPI link, or Bridge link), but of course, because it doesn't suit your purposes, you won't. I know better tho...

NUMA is Non-Unified Memory Architechture, and yes, current implementations are slow, however they are not ringbusses. HT 3.0? bandwidth?:lol2: 20.xGB?:lol2: How's 50GB, in 256-bit ringbus...oh, that's right, more than twice the bandwidth.:stick: And enough, if as you say they merely pass rendered frame data back and forth, but then, in the instance of SFR and HDR...what happens?

Gotta come up with better than that...nothing you have mentioned really prevents dual-gpu configs from sharing memory access, nor does it prevent dual-gpu configs on a single substrate. They are obstacles, sure, but they are obstacles already overcome.:fact:


However, I must admit, that maybe none of these features get used...as you say, AMD isn't known right now for been wholey upfront about thier products

16/16 vs 8/8 gives 1000 more? Sure its not due to either chipset differences or maybe texture loading times? Tried testing the same platforms with 16 vs 8 singlecards?

So you wanna compare software CF with bridge CF? Can you show the different scores for both? Or is it more like software is ok, but bridge is faster for the rendered frame merge? PCIe is not latency friendly...

With your CPU and GPu comparision. It is somewhat possible to do the same on a GPU yes. Its just impossible I/O wise. Very impossible. Also you would need quite abit bigger cache on your GPU. And whats the speed difference with the L2 vs main memory? 10x? Again, you gonna need something like 120GB/sec for the GPU part to work. Just imagine the prices and endless metallayer counts and endless amount of tracelines. Its hard enough as it is today just to get 512bit done for the memory. Also you can not compare the L2 with the framebuffer. Since the other GPU cant snoop, write or read in it.

For HT3.0 you seem to lack its unidorectional nature as an example. A 8GB/sec HT link aint 8GB/sec as such. Its 8GB/sec combined. 4Gb/sec each way and no, they cant change direction. Its like saying 100Mbit full duplex is 200Mbit. So there is suddenly a long way to the single GPUs memory bandwidth

HDR? What happens? Do you mean AFR/AFR2 or SFR?

Nothing prevents it? Well..tell me why there has been no multi GPU card yet with shared memory. yet we had them for like 10 years now in different shapes. With all this CF/SLI fuss you would expect it aswell wouldn´t you. Specially considering the GX2 card.

As I said, its a wish dream. Not impossible, just impractical on all levels.

[cadaveca]

Again, you are being too coarse. I do not intend to say ALL framebuffer is shared, and each card does not get same data..of course some data must be duplicated, but it's more than possible to render a frame, pass it off to the other gpu, and have it do an HDR pass. Or, it's possible to poll data from frames buffered in order to get lighting right, like in the instatance of SFR(Split frame rendering) in conjunction with HDR...if the lightsource is in the top frame, how does the bottom frame know where the light source is?


Now, you back track of HT3.0...I KNOW it's SLOW...but it's enough for cpu's, including quads. So, now, you take the theoretical max of 20.8gb(unidirectional, as you say), and you compare 10.4GB with 50gb of ringbus...you could, technically, situate 10 quads on that bus. As such, any minor data swaps between gpu's is not a big deal.


We aren't talking about sharing GB's of data...were talking at max 1gb of data, with 512MB cards. So, with 50gb/sec, that's the same as swapping the frambuffer completely 50 times, for both cards. A rendered frame does not weigh in @ a GB...far from it, so the end result means possible swapping of at least 100FPS, assuming 500mb/frame.

That seems like enough to me....

I don't think shared frambuffers across pci-e bus makes much sense either, but i know that with pci-e 2.0, it can be done alot better than it is now. Rough estimate says that with twice the bandwidth, your loss in performance would be at least half(in comparison to using the on-board composting engine's link) as it is now in software, which in 8x8 config, only amounts to a 15% loss in '06, and we know that '06 isn't a real app, by any means.

You seem to miss out that I'm refering to using shared memory buffers on the same board...not across pci-e, however the mention of pci-e was wholey becuase it works for crossfire now, so any link that provides as much bandwidth as the ringbus does in comparison to pci-e ensures that it's far more than possible to have a shared buffer on the same vga board. You jsut need the right interface, and from R580 on, it's there.(R580 had half ringbus disabled, was truly 512-bit).


Sry, but you went off on a tangent there, maybe by my own posts.:shrug:

[ColonelCain]

If it is true that there will be 2 GPU dies on a chip, then that seems to follow the goals for the R700.
This just might be what ATi needed to hold up against the G90 until the R700 comes out.

[madcho]

If it is true that there will be 2 GPU dies on a chip, then that seems to follow the goals for the R700.
This just might be what ATi needed to hold up against the G90 until the R700 comes out.

Ring bus is the best Way ;)

[turtle]

GPUs dont work like CPUs. CPUs can share memory. GPUs cant.

So you will need double the memory for 2 cores. Or some seriously super I/O that would kill the cost.


I didn't see this thread until now, but oddly enough I was curious about the same thing. Jawed@B3D was very insightful in explaining how it could work. Given the fact how in-depth (see 'ass-deep') this guy is in every aspect of GPU technology, I'm going to go out on a trolling and say you're probably wrong.

Have a nice day.



Me:

If you had two 256-bit dies put together [on] one package sharing pins etc, could they share a 512-bit bus, or could a 256x2 configuration work with some kind of load balancing beyond what we see in current crossfire?

Jawed:

The 256-bit bus on most GPUs is actually a collection of smaller buses, e.g. 4x64-bit or 8x32-bit.

What I'm expecting to see is that two GPUs on one package have a fat bus joining them (much like Xenos and its daughter die have a 32GB/s bus). The two separate memory systems (256/512MB per GPU) are then aggregated into a single memory address space.

In effect GPU A can request a texture from GPU B's memory, so the fetch request travels over to the relevant memory controller on GPU B which then obtains the data GPU A requires. All clients in both GPUs (TUs, RBEs, vertex fetch, etc.) see this single memory space and are able to use it freely.

Then you just need a decent driver that understands how best to assign memory to the clients on both GPUs, so that the chosen multi-GPU mode produces the most efficient usage of memory as well as the required performance.

AFR seems like the prime candidate. Within AFR, though, it's possible to optimise the way textures are organised - e.g. classically textures are copied to both GPUs' memory. In theory the newer ATI GPUs don't need to be that wasteful.

I've got my fingers crossed that we'll see these kinds of efficiency gains, but the driver gods have been scowling upon these D3D10 GPUs and I see no sign of a let up.

Jawed

You can see the whole conversation here (http://forum.beyond3d.com/showthread.php?t=42711&page=17), and the gist I get is that it is not only possible, but likely.

[Shintai]

I didn't see this thread until now, but oddly enough I was curious about the same thing. Jawed@B3D was very insightful in explaining how it could work. Given the fact how in-depth (see 'ass-deep') this guy is in every aspect of GPU technology, I'm going to go out on a trolling and say you're probably wrong.

Have a nice day.



You can see the whole conversation here (http://forum.beyond3d.com/showthread.php?t=42711&page=17), and the gist I get is that it is not only possible, but likely.

If you actually took your time to read what i wrote. I didnt claim it was impossible, but rather completely impractical due to I/O limitations (Read fat bus). You gonna need something like 120GB/sec full duplex or more in reality. While the internal I/O in the GPU also needs close to a doubling. Along with being very cheap and dont increase PCB cost alot. It just dont mix.

And the Xenos eDram sits on a memory controller. Its not an intercommunication I/O. And its half duplex. So even then its about 7-8x short in performance. And thats the continual problem we had since the Voodoo and MAXX days.

As I said, its a wish dream. Not impossible, just impractical on all levels.

The sole reason to use 2 dies on a package would be yield of the dies. But again it would add cost plus the I/O parts of a GPU to another GPU. It would be much cheaper just to make the larger single die. Since defects can usually be recovered with disabled pipe(s).

Because its possible doesnt mean its economic or practical realistic.

[Face]

http://my.ocworkbench.com/bbs/showthread.php?%20threadid=67962 (http://my.ocworkbench.com/bbs/showthread.php?%20threadid=67962)

bluetooth
Location: 中华江南
Posts: 49,639

AMD R680 can't even reach 18600 in 3Dmark06
--------------------------------------------------------------------------------

There are some rumours on the net claiming that he R680 will score or rather aim for 20,000 points in 3Dmark06.

From our source, it is said that R680 won't even get beyond the 18600 mark. Speculations and rumours goes around whenever there is an upcoming product.

Anyway, the slides that were shown were dated and does not reflect the final specs. Remember, specs always changes without prior notice.__________________

[Eastcoasthandle]

http://my.ocworkbench.com/bbs/showthread.php?%20threadid=67962 (http://my.ocworkbench.com/bbs/showthread.php?%20threadid=67962)

Source?
I read the link and there is no real source for this information to validate it.

[Zytek_Fan]

Source?
I read the link and there is no real source for this information to validate it.

Now that R680 is up for debate the rumours are going to be thrown back and forth...

[SunFlowerSeeds]

Nooo we dont want another R600. Please AMD...

[Face]

Source?
I read the link and there is no real source for this information to validate it.

OCWorkbench _is_ the source. bluetooth is the one writing the news. its on the front page http://ocworkbench.com/ or at http://www.ocworkbench.com/news/news.php it's just appears as a forum post.

They are very reliable AFAIK.

Ofcourse that it's all rumors at this point, so everything is a bit shaky and naturally nothing is written in stone.

[BrowncoatGR]

OCWorkbench _is_ the source. bluetooth is the one writing the news. its on the front page http://ocworkbench.com/ or at http://www.ocworkbench.com/news/news.php it's just appears as a forum post.

They are very reliable AFAIK.

Ofcourse that it's all rumors at this point, so everything is a bit shaky and naturally nothing is written in stone.

So is VRzone. I think its better to wait until more info gets released or even better until some XS member posts some scores

[Syn.]

I'm excited for R680, but I don't want to get my hopes up too much. I want an all AMD system that rocks :(

man... i just want a decent DX10 card from AMD this year, is that too much to ask for these days?


Also dont care about nvidia, being a vista user.

[adamsleath]

i keep hearing ati has bad drivers for AA&dx10 vs nvidia 8800's...
anyone got any more info on this? like an example maybe?