Samsung begins shipping GDDR4 modules...

[DilTech]

http://www.eetimes.com/news/semi/showArticle.jhtml?articleID=190300221


Samsung's 512-megabit GDDR4 memory will be offered in a 32-bit data bus configuration.

"With GDDR4-based systems, we're entering a new frontier with life-like movements and richly textured imagery far better than DVD movies being viewed today," said Mueez Deen, marketing director for Samsung Semiconductor, based here, in a statement. "Using 80-nanometer production technology, the memory has an ultra-high-speed data processing rate of 2.4Gigabits per second (Gbps)."


It has begun...

[Lightman]

Best part is lower POWER USAGE!!

[Rovtar]

who will used it first?

[Mr. Popo]

Yay! :woot:
So now that GDDR4 is old news, what about XDR? :D

[Cooper]

Modules ? Ir just chips ? :cool:

[Pinnacle]

DO we have a DDR4 guru?

[gOJDO]

who will used it first?

graphics cards of course :)
Modern graphics cards with 512bit memory interface can have 1GB of GDDR4 memory with total bandwidth of 153.6GB/s(2.4Gbps * 512bits / 8bits). It will operate at 2.4GHz.
GDDR4 means that we will see DDR3 memory for mainboards soon.

[Cooper]

One year to come sounds soon enough :D
Soon GPU vendors will compete not only with memory size, but speed will increase also. Current chips are B/W sturving.

[nn_step]

DO we have a DDR4 guru?
Actually GDDR4 is closer to DDR3 than anything ;)
low power, high bandwidth and good latency. what isn't to like?

[Cooper]

Any link to specs ?

[Lightman]

who will used it first?
If You asking about which company I think ATI on R580+ cards

[lookmomnobrains]

who will used it first?

nvidia on the first gen dx10 cards followd 6-8 weeks later by ati...

[Lightman]

graphics cards of course :)
Modern graphics cards with 512bit memory interface can have 1GB of GDDR4 memory with total bandwidth of 153.6GB/s(2.4Gbps * 512bits / 8bits). It will operate at 2.4GHz.
GDDR4 means that we will see DDR3 memory for mainboards soon.

something is fishy!

How on earth we can get 153.6GB/s transfer rate from memory when processor cache is hitting only XXGB/s??
Could anyone provide link to GDDR4 spec sheets??

[Lightman]

nvidia on the first gen dx10 cards followd 6-8 weeks later by ati...

:nono:
before nVidia G80 Ati will relelase R580+ with GDDR4

[largon]

Availability will be the biggest obstacle for the usage of these chip since GDDR4 is identical compared to GDDR3 for what comes to I/O stream, thus GPU mfg's could slap them on current cards.
Though, it's likely they won't be doing that at least for some time since it's likely that current GPU MC's would form the biggest bottleneck as they'r not designed to handle the frequencys these chips run at. That is, of course unless some GPU mfg's already have an MC that scales well enough...

:hehe:


0.6ns (3.2GHz effective):

[nn_step]

god those are some sexy chips

[gOJDO]

If You asking about which company I think ATI on R580+ cards
There was BS about it, as usual from the inq( http://www.theinquirer.net/default.aspx?article=31864 ), about ATi using DDR4 in R580+, but actually nVIDIA were those who were involved in the development:
"NVIDIA worked with JEDEC to develop the GDDR4 standard and we are excited to see Samsung ramping this into production," said Tony Tamasi, vice president of technical marketing at NVIDIA. "The graphics industry has an insatiable demand for high bandwidth memories and GDDR4 is poised to become the next major evolutionary step for our GPUs."
ATi hope that they can use it:
"We're delighted that we'll be able to use GDDR4 from Samsung in our latest graphics cards," said Joe Macri, senior director of engineering, ATI, and chair of the JEDEC committee on GDDR4. "Samsung's timely introduction will increase the performance of our upcoming products and ultimately improve the gaming experience for our users".

http://www.samsung.com/us/Products/Semiconductor/USNews/GraphicsMemory/GraphicsMemory_20060705_0000268897.asp

[largon]

Modern graphics cards with 512bit memory interface can have 1GB of GDDR4 memory with total bandwidth of 153.6GB/s(2.4Gbps * 512bits / 8bits). It will operate at 2.4GHz.
GDDR4 means that we will see DDR3 memory for mainboards soon.Eh, 512bit interface? No such thing out there.
It would require like 10+ layer pcb and 16 ram chips.
:confused:

Let's do a little comparison:

1.
R580 X1900XTX w/ GDDR3 @ 0.775GHz (1.55GHz):
1.55Gbps * 256bits / 8bits = 49.6GB/s
775MHz = "cream of the GDDR3 crop"

2.
Low bin GDDR4 @ 1.2GHz (2.4GHz):
2.4Gbps * 256bits / 8bits = 76.8GB/s (+55%)
a low speed for GDDR4...

3.
Moderate bin GDDR4 @ 1.8GHz (3.6GHz):
3.6Gbps * 256bits / 8bits = 115.2GB/s (+132%)

[K404]

low power, high bandwidth and good latency. what isn't to like?

...having to wait for it? :p:

[gOJDO]

Eh, 512bit interface? No such thing out there.
no, a lot of out there.
It would require like 10+ layer pcb
no.
and 16 ram chips.
yes.
:confused:
no need to.

Let's do a little comparison:

1.
R580 X1900XTX w/ GDDR3 @ 0.775GHz (1.55GHz):
1.55Gbps * 256bits / 8bits = 49.6GB/s
775MHz = "cream of the GDDR3 crop"

2.
Low bin GDDR4 @ 1.2GHz (2.4GHz):
2.4Gbps * 256bits / 8bits = 76.8GB/s (+55%)
a low speed for GDDR4...

3.
Moderate bin GDDR4 @ 1.8GHz (3.6GHz):
3.6Gbps * 256bits / 8bits = 115.2GB/s (+132%)

Right, DDR4 will scale to 4.8GHz and with 512bit interface it will provide 307.2GB/s. (+520%)

[nn_step]

Right, DDR4 will scale to 4.8GHz and with 512bit interface it will provide 307.2GB/s. (+520%)
:slapass: IT IS GDDR4 goddammit :slapass:
Say it right

[lookmomnobrains]

:nono:
before nVidia G80 Ati will relelase R580+ with GDDR4

ATI R580+ is expected to be released in August / September while g80 is expected to be released in September / October so its still to to early to tell ho gone be first.. not that that madders really much.. :fact: but looking at them dates your probably rite :buddies:

[largon]

gOJDO,
Ah, my mistake.
(Top of the line ATi FireGL (http://www.ati.com/products/fireglv7350/index.html) cards have 512bit memory)

[[XC] Lead Head]

THERE are no GPUs on the market with 512bit memory interface, the r580/520 have a memory controller CAPABLE of 512bit, but their pin outs are only for 256bit. The G70/71 is 256bit internal and external

[gOJDO]

THERE are no GPUs on the market with 512bit memory interface, the r580/520 have a memory controller CAPABLE of 512bit, but their pin outs are only for 256bit. The G70/71 is 256bit internal and external
GeForce 7950 GX2
Graphics Bus Technology PCI Express
Memory 1GB (512MB per GPU)
Memory Interface 512-bit
Memory Bandwidth (GB/sec) 76.8
Fill Rate (Billion pixels/sec) 24
Vertices/second (Billion) 2.0
Pixels per clock (peak) 48
RAMDACs (MHz) 400
http://www.nvidia.com/page/geforce_7950.html

Industry’s First Ultra High-End 1GB Workstation Graphics Accelerator
Introducing the ATI FireGL V7350 with Avivo™ Technology – the ultimate graphics accelerator designed for the most complicated 3D models, the largest data sets, and highest definition textures. The FireGL V7350 delivers industry leading features and performance for the most demanding workstation users running OpenGL and DirectX based applications.
Features
Powered by ATI’s next generation FireGL™ graphics processor unit (GPU) with Avivo™ Technology
Scalable ultra-threaded architecture with 8 Parallel Geometry Engines and 16 Pixel Shader Processors
Full Shader Model 3.0 support for vertex and pixel shaders
1GB GDDR3 graphics memory with 512-bit ring bus memory controller
High Dynamic Range (HDR) rendering with 8-bit, 10-bit, and 16-bit per RGB color component support
High fidelity display engine capable of producing over one trillion colors
Two Dual link outputs each capable of ultra-high resolutions up to 3840 x 2400
Stereoscopic 3D output
HD Component Video (YPrPb) output
Optimized and certified for CAD and DCC applications
Direct access to ATI’s dedicated workstation technical support team
http://www.ati.com/products/fireglv7350/index.html

[flutie98]

http://www.nvidia.com/page/geforce_7950.html


http://www.ati.com/products/fireglv7350/index.html



iirc 512bit ring bus is 256 bit bi direcitonal, stupid marketing gimmic

[Vapor]

Yeah, and the 7950's 512-bit is really just two 256-bit interfaces (one per GPU).

Regardless....there's no SIGN of ATi/nV using real 512-bit interfaces to the memory any time soon on consumer cards, so we're at 76.8GB/s with 2.4GHz GDDR4.

Later GPUs and higher memory speeds may be accompanied by a 512-bit interface, but that's way later.

[Cobalt]

The data need to go both ways so as far as performance is concerned (and the amount of traces etc) it is effectivly 512bit internally.

Besides, we've had 256bit for ages, its about time they upped it, even if it isn't going to happen till nex year (realistically)

[flutie98]

a true 512bit interface is a emc nightmare right now, which is why weve been using 256bit since the 9700 years

[gOJDO]

No metter how, the point of graphics card is to boost graphics performance. It is somehow achieved with two cores on one graphics card. Two cores means twice memory bandwidth and twice memory capacity. With the same memory interface GDDR4 will offer 33%-166% better performance than GDDR3. But, as one of the discutants mentioned, it is just a question of time when we are going to see mulitcore graphics chips with widther memory interfaces on one graphics card.

[[XC] Lead Head]

No metter how, the point of graphics card is to boost graphics performance. It is somehow achieved with two cores on one graphics card. Two cores means twice memory bandwidth and twice memory capacity. With the same memory interface GDDR4 will offer 33%-166% better performance than GDDR3. But, as one of the discutants mentioned, it is just a question of time when we are going to see mulitcore graphics chips with widther memory interfaces on one graphics card.

No. The 7950GX2 is two seperate PCBs, each core with its own memory interface, the total efective memory capacity and bandwith is not doubled.

And like I said earlier, the r520/580 is 512 bit INTERNAL and 256bit EXTERNAL.

ps.if your avatar is what i think it is, thats grounds for you getting banned.

[gOJDO]

No. The 7950GX2 is two seperate PCBs, each core with its own memory interface, the total efective memory capacity and bandwith is not doubled.
The effective capacity is ike 512MB, but the bandwidth is twice.

P.S. Avatar removed.

[[XC] Lead Head]

The effective capacity is ike 512MB, but the bandwidth is twice.

P.S. Avatar removed.

I don't get that logic, each core has 512MB and ~35-40 GB/s bandwith to work with. Core 1 cant access Core 2 memory and Core 2 cant access Core 1 memory

[gOJDO]

I don't get that logic, each core has 512MB and ~35-40 GB/s bandwith to work with. Core 1 cant access Core 2 memory and Core 2 cant access Core 1 memory
both cores are working on the same task, but both cores have own textures cache copy.

[[XC] Lead Head]

both cores are working on the same task, but both cores have own textures cache copy.

No. Alternate frame rendeing, Core 1 renders even number frams, Core 2 renders the odd, can have completely diffrenent textures perfram, in scissor mode, each core renders half the screen, different textures again, and checkerboard where Core 1 renders the odd squares and Core 2 renders even, once again, different textures

[gOJDO]

No. Alternate frame rendeing, Core 1 renders even number frams, Core 2 renders the odd, can have completely diffrenent textures perfram, in scissor mode, each core renders half the screen, different textures again, and checkerboard where Core 1 renders the odd squares and Core 2 renders even, once again, different textures
That is what I said exactly, they have their own copy. It is like a core with twice hardware potentionals with 512MB memory for textures and other data, but with twice the bandwidth.

[nn_step]

That is what I said exactly, they have their own copy. It is like a core with twice hardware potentionals with 512MB memory for textures and other data, but with twice the bandwidth.
Dude stop digging yourself into a hole :rolleyes:

[Stuperman]

THERE are no GPUs on the market with 512bit memory interface, the r580/520 have a memory controller CAPABLE of 512bit, but their pin outs are only for 256bit. The G70/71 is 256bit internal and external

wildcat realism 800 uses 512-bit memory.

[perkam]

G80 most likely first use, R600 second.

Perkam

[[XC] Lead Head]

wildcat realism 800 uses 512-bit memory.

3D labs is leaving the GPU market

[gOJDO]

Dude stop digging yourself into a hole :rolleyes:
Dude, stop being ignorant and blind and stop arguing. You have no abilities to support any argument you make and you missunderstood most of the links you are providing as proves. I will suggest you to stick at the few things you know right.

http://techreport.com/reviews/2006q2/geforce-7950-gx2/block-diagram.gif
http://techreport.com/reviews/2006q2/geforce-7950-gx2/index.x?pg=2
Because this is SLI on a card, it has some limitations. You may see the 7950 GX2 advertised as a 1GB card, and it undeniably has that much video RAM onboard. Yet that RAM is segregated into two 512MB pools, one for each GPU. Yes, the GX2 has about twice the memory bandwidth of a normal card, but functionally, it has a 512MB memory space. Textures and other data must be uploaded to each GPU and stored in each GPU's associated memory.

http://www.gamepc.com/labs/view_content.asp?id=7950gx2&page=2

. Thus, in effect, a single GeForce 7950 GX2 card has 48 rendering pipelines (24 per GPU), 16 shader pipelines (8 per GPU) and can provide double the memory bandwidth, as each GPU has a 256-bit memory interface. Even with this in mind, we still don't feel entirely comfortable saying that the GeForce 7950 GX2 cards have 48 pixel pipelines with 512-bit memory interface, as this still is somewhat misleading. 2 x 24 pipeline and 2 x 256-bit interfaces would be the proper terminology, which we're going to stick with for the time being.

Read the specifications about the card on the table at the end of the page:
http://www.prolink.com.tw/english/products/vga/PCX/GeForce%207950%20GX2.htm

[[XC] Lead Head]

NOTHING IS DOUBLED. That is marketing gimmic!!! Thats like saying i have two 7900GTs in SLI it has ~70-80GB/s mem bandwith, 48 shaders, etc.. when it simply isnt true!

[alexio]

We won't see 512bit soon. 256bit 1.2ghz+ is already very hard to design an MC and PCB for. Besides that, next-gen GPU's will not need that kind of bandwidth. It's better to spend more money on producing a bigger GPU or dual-die videocard than to spend tons on a 512bit capable MC and PCB.

Dual-GPU 256bit*2 we will most likely see before 512bit single GPU on next-gen cards.

[gOJDO]

NOTHING IS DOUBLED. That is marketing gimmic!!! Thats like saying i have two 7900GTs in SLI it has ~70-80GB/s mem bandwith, 48 shaders, etc.. when it simply isnt true!
Hey, calmn down. The performance scalling of the SLI is not 100% efficient, but there is performance improvement compared to singlecore graphics card. And that is the goal. With less clocked core and memory, the perofmance of the two cores graphics card is better and in reasonable amount.
How do you explain the performance boost?
http://www.pcper.com/images/reviews/256/cod2-48-bar.gif
http://www.pcper.com/images/reviews/256/fear-48-bar.gif
http://www.pcper.com/images/reviews/256/bf2-hd-bar.gif
http://www.pcper.com/images/reviews/256/3dmar06-1280.gif
http://www.pcper.com/article.php?aid=256

[[XC] Lead Head]

Hey, calmn down. The performance scalling of the SLI is not 100% efficient, but there is performance improvement compared to singlecore graphics card. And that is the goal. With less clocked core and memory, the perofmance of the two cores graphics card is better and in reasonable amount.
How do you explain the performance boost?
http://www.pcper.com/images/reviews/256/cod2-48-bar.gif
http://www.pcper.com/images/reviews/256/fear-48-bar.gif
http://www.pcper.com/images/reviews/256/bf2-hd-bar.gif
http://www.pcper.com/images/reviews/256/3dmar06-1280.gif
http://www.pcper.com/article.php?aid=256

Because the cards each get their own frame to render....:stick:

[zakelwe]

The GX2 does not have twice the memory bandwidth because the framebuffer is replicated for both so in effect it has the same memory bandwidth per chip. Bandwidth value here should be per chip, not per board.

7900GT = 41.6GB/sec
7950 GX2 = 2x38.4GB/sec

Note it is 2x38.4 and NOT 76.8 !


In regards to general desktop gpu 512bit I doubt you will see it whilst memory performance keeps increasing as the cost and techniques for manufacture are more than just slapping on faster memory which gets cheaper with time. GDDR4 will put back 512bit even further into the future, if ever.




Regards

Andy

[gOJDO]

Because the cards each get their own frame to render....:stick:
you mean each chip on the card. ;)
exactly, they have their own frames to render. Each G71 chip on the 7950GX2 independently renders N number of frames. Becouse there are two chips(twice hardware potentional) it has twice hardware requirements(twice memory), but their efficiency is not 100%. It has 512bit memory interface, 256bit for each core, but each core can only use its own 512MB as memory resource. The cores are not sharing the resources, they are sharing only the graphics card tasks.
The memory bandwidth on graphics card level is doubled. The two cores are handling twice bandwidth for rendering.

[Cobalt]

Why is everyone beating on gOJDO? He's right you know and everyone is saying the same thing. I think there was some misinterpretation of a post early on and now you're all attacking what he says.

All he said was that the GX2 does not have a true 512bit interface or 1GB VRAM. Now he had to cite all that evidence for something that everyone else knew all along :rolleyes:

[[XC] Lead Head]

you mean each chip on the card. ;)
exactly, they have their own frames to render. Each G71 chip on the 7950GX2 independently renders N number of frames. Becouse there are two chips(twice hardware potentional) it has twice hardware requirements(twice memory), but their efficiency is not 100%. It has 512bit memory interface, 256bit for each core, but each core can only use its own 512MB as memory resource. The cores are not sharing the resources, they are sharing only the graphics card tasks.
The memory bandwidth on graphics card level is doubled. The two cores are handling twice bandwidth for rendering.

No. The 7950 is two graphics cards, bolted together and SLIed. Its not efectively doubled.

[lm358]

Jeez lets just say you're BOTH right, depending on your classification of the 7950? :-)

[gOJDO]

No. The 7950 is two graphics cards, bolted together and SLIed. Its not efectively doubled.

It is one graphics card, with two boards connected with advanced 48 lines. It is represented as a single card and it is dividing tasks automaticly. Off course it is not effectively doubled, but consider the clock reduction of the graphics cores and the video memory and the handicap of PCIx16 while feeding two graphics chips instead of one. 7950GX2 singlecard works on non SLI mainboards. By using SLI, with two 7950GX2, QUAD SLI is done, just like DUAL SLI is done with other singlecore graphics cards. And the performance is again better than singlecore cards SLI configuration.

[_damien_]

I agree with you, gOJDO.


A couple of thoughts:

Everybody seems to agree that SLI is a pretty efficient method of task-sharing (which isn't always true for dualcore CPUs, considering the software-part)

The GPUs share the same task, have a dedicated memory pool / bus, and the rendering capacity is doubled (in theory). So considering that all the bandwidth is available to the same task, I think it makes sense to state that the bandwidth is actually doubled.


And if it's not, please consider the following...


* Nobody claims that the bandwidth of an AMD X2 is cut in half compared to single cores. (Neither do I, only stating this for the sake of argument):

The cores share the same memory subsystem, but the perception is that the bandwidth remains the same, right? Following the logic of certain forum members, how is the BW not cut in half (Athlon X2), if it remains the same even when adding a new core that has its dedicated controller/memory (7950X2) ?

* Did anybody ever dispute the fact that a 2-way, NUMA-aware Opteron-system has double bandwidth compared to a 1-way system?

Not that I've heard. And the concept is basically the same as SLI / 7950X2, right? In both cases there are two cores, two controllers, two buses and two memory pools. In my book that implies doubled bandwidth.

[Cobalt]

The think is that its only double *total* bandwidth. In the case that one card requires more bandwidth it can't dig into the supply of bandwidth from the other. Thats where load balancing comes into play but its still the essential reason why its not *true* double bandwidth.

Not that it matters to me. I couldn't care less what the specs are as long as it kicks ass in the performance side. i.e. where it counts ;)

[[XC] Lead Head]

The think is that its only double *total* bandwidth. In the case that one card requires more bandwidth it can dig into the supply of bandwidth from the other. Thats where load balancing comes into play but its still the essential reason why its not *true* double bandwidth.

Not that it matters to me. I couldn't care less what the specs are as long as it kicks ass in the performance side. i.e. where it counts ;)

If you mean memory bandwith, no. Each core can only access its own memory bank

[Cobalt]

Ooops that was a typo. I ment to write "can't" :doh:

[gOJDO]

you can think whatever you wish, but it is twice the bandwidth.

[Cobalt]

Does everybody in this thread intentionally misunderstand what everybody else posts? I really can't be bothered to continue to post here.

[largon]

you can think whatever you wish, but it is twice the bandwidth.Sure, one could also say 79_0GX2 has twice the GPU frequency compared to a traditional card. Or that two identical computers connected via ethernet combined have double cpu frequency, memory, etc. One can say so, but in the end it doesn't mean anything relevant.

This discussion is just about semantics.

[gOJDO]

no, but can say it has twice rendering & shader pipelines.
when you say that singlecore card has 256bit memory interface, you are not saying it has 8x32bit becouse 8 chips are involved.
Also don't confuse about one thing, the total bandwidth is theoretical and it can be achieved only in the case when the graphics chip(s) are handling all the memory chips. Each chip provides freqfency*data width/8 bytes/s, for example the GDDR3 cihps are 32bit and at 1200MHz each will provide 4.8GB/s, multiplied to 8 it is 38.4GB/s. But if the data needed for rendering by one graphics chip is stored on one memory chip, the total theoretical bandwidth will not be 38.4, it will be 4.8GB/s. And there is no specification with 8x4.8GB/s, but it is 38.4GB/s. When two grpahics cores are involved, the bandwidth is doubled becouse there are 16 memory chips, each 32bit and each providing 4.8GB/s or total 512bit memory interface with 76.8GB/s total bandwidth.

[zakelwe]

you can think whatever you wish, but it is twice the bandwidth.

NOT IT IS NOT.

It is the same bandwidth but done twice over because there are two gpu's on the board.

Each gpu is getting x1 bandwidth and not x2. The workload on bandwidth is not split because the framebuffer is not split, it is replicated on both memory systems.

http://www.beyond3d.com/misc/chipcomp/?view=boarddetails&id=319

38.4GB/sec ... that is not twice the bandwidth for 256bit memory working at 600Mhz. 256 / 8 x 1200Mhz = 38.4. If it was twice the bandwidth it would be 512 / 8 x 1200 = 76.4.

[zakelwe]

When two grpahics cores are involved, the bandwidth is doubled becouse there are 16 memory chips, each 32bit and each providing 4.8GB/s or total 512bit memory interface with 76.8GB/s total bandwidth.

NO ... from nvidia

"Is the frame-buffer memory shared, i.e., can I access a 512MB frame-buffer if I have two GPUs with 256MB each?
No, each GPU maintains its own frame-buffer. Rendering data, such as texture and geometry information, is duplicated. "


http://www.hwspirit.com/reviews.php?read=14&page=3

Although nvidia is talking about the size of the framebuffer here the important thing is that they are saying it is duplicated. Because it is duplicated then you have NOT got a 512bit memory interface and you have not got 76.8 GB/s bandwidth. Both those need dividing by two.

When we talk about bandwidth we talk per gpu, that is the important thing, not per card. Two SLi cards in a system do not mean that system has twice the bandwidth as the framebuffer is once again duplicated. Only the gpu's have ( in theory ) half as much work to do when you double them up.

Regards

Andy

[gOJDO]

NO ... from nvidia

"Is the frame-buffer memory shared, i.e., can I access a 512MB frame-buffer if I have two GPUs with 256MB each?
No, each GPU maintains its own frame-buffer. Rendering data, such as texture and geometry information, is duplicated. "
Exactly, if each GPU has 256MB then the effective memory is 256MB, the bandwidth is twice. The rendering data is duplicated and there are 2 GPUs handling its own copy=>2X bandwidth(on card, not on GPU). On card level, very simple: twice memory chips and twice GPUs to have twice bandwidth and twice rendering potentional(but not twice memory capacity).


http://www.hwspirit.com/reviews.php?read=14&page=3

Although nvidia is talking about the size of the framebuffer here the important thing is that they are saying it is duplicated. Because it is duplicated then you have NOT got a 512bit memory interface and you have not got 76.8 GB/s bandwidth. Both those need dividing by two.
Memory interface on the card(7950GX2) is 512bit, 16 chips, 32bit each 16x32=512bit. Total bandwidth is 76.8GB/s using 1200MHz GDDR3.

When we talk about bandwidth we talk per gpu, that is the important thing, not per card. Two SLi cards in a system do not mean that system has twice the bandwidth as the framebuffer is once again duplicated. Only the gpu's have ( in theory ) half as much work to do when you double them up.

Regards

Andy
Remember, 7950GX2 is dual GPU card, but it is single graphics card and it is represented as it is, not as SLI configuration of two graphics cards. It works on non SLI mainboards. Internaly it is using two GPUs and has 512bit memory interface. Two GPUs are rendering and each owns a memory controller and memory with a copy of the required data for rendering. The two GPUs are performing memory operations in paralel and they are handling twice bandwidth.

[zakelwe]

The two GPUs are performing memory operations in paralel and they are handling twice bandwidth.

They are not handling twice the bandwidth. There is 76.8GB/s bandwidth in total on the card but it is shared between two gpu's so that is the same bandwidth as if there was one gpu.

The only people who are claiming 76.8GB/s and 512bit are the graphics card companies and nvidia cause that marketing sells cards.

When you are running 2 gpu's say in SFR mode then maybe ideally they halve the workload for gpu, fine. But in regards to memory the entire screen is replicated in both memory subsystems, not half so you are not halving the memory load. You are having to work on two sets of the screen and therefore each screen it is being manipulated at a max of 38.4Gb/s. If you start upping the screen resolution and AA/AF modes then you will become bandwidth limited not at 78.6GB/s but at 38.4GB/s unfortunately. So in effect it is the 38.4GB/s that is your effectual bandwidth.

If you do not believe me stick a post up at Beyond3d.com forum and see what they say ....

Regards

Andy

[zakelwe]

Memory interface on the card(7950GX2) is 512bit, 16 chips, 32bit each 16x32=512bit.



No, the 7950GX2 uses 2 x G71 chips. The G71 chip has a 256bit memory interface.

http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=112

On the 7950GX2 there are 2 x 256 bit memory interfaces and this is not the same as saying one 512bit memory interface. You keep doubling things then saying that there is twice the bandwidth or twice the bus width .. this is not correct unless you are MSI or Gainward and want to have big numbers in your marketing gumf.

Regards

Andy

[gOJDO]

ok.

[krille]

Actually GDDR4 is closer to DDR3 than anything ;)
low power, high bandwidth and good latency. what isn't to like?Incorrect. We've been over this before nn_step, you should listen for one's sake. GDDR4 is tweaked GDDR3 is tweaked GDDR2 is tweaked DDR2. It is not in any way based on DDR3, DDR3 is another elaboration of DDR2, another branch if you wish. DDR3 is something completely different from GDDR3 and GDDR4. Don't confuse GDDR with DDR. Although they can admittedly share some characteristics (lower power consumption etc) as lessons learned on one branch often is applied on the rest as well. :rolleyes:

gOJDO just calm down and go make a poll or something. If people haven't agreed with you so far, I doubt they will ten pages from now either.

~ Kris