Sign in with
Sign up | Sign in
Your question
Solved

384-bit memory bus AMD/ATi

Last response: in Graphics & Displays
Share
Anonymous
a b U Graphics card
a b } Memory
March 21, 2011 4:23:14 PM

Hi,
Ati was the first to use GDDR5 on it's graphics cards I think,
but they still use 256 bit memory interface where NVIDIA Uses 320/384-bit memory interface at its high-end.
Please tell me why AMD seems to forget about this very important factor.
Another question is that when they say 512-bit memory interface (2x256 bit) for cards like 6990 is it right that we ADD two interfaces together? the efficiency of 2x256 lane is equal to 512?
Thanks.
a c 235 U Graphics card
a b } Memory
March 21, 2011 4:39:33 PM

because its not an important factor. nvidia and AMD engineer their cards differently. sometimes they use the bus to slow a card down in order to meet a market segment, sometimes increasing the bus will show no improvement at all so there is no point.

the AMD has two gpu cores, each core has a bus width of 256. there is not a single core with 512 in the 6990.
Anonymous
a b U Graphics card
a b } Memory
March 21, 2011 4:44:49 PM

It is obvious that 6990 has 2 GPUs anyway. but I dont think lowering the bus is better at all.
and you did not answer my second Q. ct1615 ;) 
Related resources
a c 235 U Graphics card
a b } Memory
March 21, 2011 4:46:50 PM

what was your second question? as far as what you think about lowering bus, well its pointless (you are not AMD) and seems to be incorrect based on your stated opinions.
Anonymous
a b U Graphics card
a b } Memory
March 21, 2011 4:51:35 PM

My question is clear ct... .Pardon me but You were not forced to reply, thank you.
a b U Graphics card
March 21, 2011 5:07:14 PM

Total memory bandwidth is what is important, it doesn't matter whether you achieve it with higher clock speed and a narrower bus or slower clocks and a wider bus.

Best solution

a c 179 U Graphics card
a b } Memory
a b À AMD
March 21, 2011 7:19:33 PM
Share

A wider memory bus means more expensive board design, as you have more traces on the board, and a bigger memory controller. They all play with a variety of designs and pick the best one, if moving up to a 384 bit memory interface only adds 2% more performance in most things but will cost 10% more its a poor choice.

Memory bandwidth is not the limiting factor in the majority of graphics situations, only when you begin turning up AA at high resolutions does memory capacity and bandwidth become a limiting issue.


As for 2x256, you cannot treat it as a single 512-bit bus because each 256 bit bus feeds a separate GPU, and they both get fed pretty much the same data because their memory stores pretty much the same data just like having two cards in crossfire, they just share power regulation circuitry and board space, but the GPUs and memory are isolated systems for each.
a c 172 U Graphics card
March 21, 2011 9:06:50 PM

ATI only did one mega chip, 420mm2 and 512bit physical path. Nvidia is able to get away with mega chips to their designs and large dies. ATI typically keeps their high end card around 350mm2+ and nothing more with a 256bit. They could do a mega chip however you can kiss high clocks and efficiency good by. Over all this is only academic.
Anonymous
a b U Graphics card
a b } Memory
March 22, 2011 4:54:04 AM

Quote:
As for 2x256, you cannot treat it as a single 512-bit bus because each 256 bit bus feeds a separate GPU, and they both get fed pretty much the same data because their memory stores pretty much the same data just like having two cards in crossfire, they just share power regulation circuitry and board space, but the GPUs and memory are isolated systems for each.


Yes. in my opinion that must be this like not the text on some Card's Packages says:512 bit interface or so on.
thanks a lot hunter & and you nforce. more information about why AMD is behind would be appreciated.
:sarcastic: 
a c 235 U Graphics card
a b } Memory
March 22, 2011 1:49:55 PM

AMD is not behind, where do you get this crap from? I'm an Nvidia fan but I would say the AMD 5xxx and 6xxx series have been better to much better then Nvidia 4xxx & 5xxx. If anything its Nvidia that has been behind.
a c 172 U Graphics card
March 22, 2011 2:57:33 PM

You should see just how terrible the trolling has gotten on youtube in the past week or so. :o 
a c 130 U Graphics card
a b } Memory
March 22, 2011 4:46:33 PM


I have noticed that also but like this poster says we don't have to reply. there have been a few posts now where i just thought. yea right then son you just get on withit if you know better :pfff: 

Anyway some facts about the 256 bit bus and 256 bit bus situation.

GDDR5 runs in basic layman's terms Quad pumped while GDDR4 and3 run double pumped, so its basically shifting the same amount over the bus as would GDDR3/4 over a 512 bus anyway.
Thats basically why they don't need to use a bigger bus, also as said its cheaper to use a smaller bus. That and there is a trend in modern games to put more load for want of a better word on the engine rather than the bus.

Mactronix :) 
a b U Graphics card
March 22, 2011 5:00:12 PM


agreed, I appreciate the curiousity but likely this info will go to someone that will never own a 6990 or anything close to it....

OP: do not worry about 256 bit vs 320 bit vs etc, they are all engineered differently and the main thing you want to look at is the total video throughput and what performance it gets in various benchmarks.... do not worry about specs so much a gpus performance is all relative to a hierarchical listing of other comparable cards, for example the 256 bit 6970 is way faster than the 320 bit gtx 470.... this example right here makes your point moot, amd is not behind by any means, both companies are on the cutting edge of technology as far as 3d graphics are concerned...
a b U Graphics card
March 22, 2011 5:09:28 PM

I keep thinking some community college ESL class is giving out assignments to start an argument on a message board to practice your english.
a b U Graphics card
March 22, 2011 5:10:26 PM

^^yeah broke as hell to say the least
a c 235 U Graphics card
a b } Memory
March 22, 2011 5:19:51 PM

benski said:
I keep thinking some community college ESL class is giving out assignments to start an argument on a message board to practice your english.


unfortunately trying to read their posts is like trying to read a text conversation between two thirteen year old girls at the mall who just saw justin bieber....
a b U Graphics card
March 22, 2011 5:21:38 PM

^^^GRRR Bieber...I hate his "wannabe ghetto" little accent
Anonymous
a b U Graphics card
a b } Memory
March 24, 2011 5:41:26 AM

Thank you guys even if I did not agreed with some of your posts.
Note that when I said "AMD is behind I meant 256 is smaller than 384" and if AMD is edge to edge with NVIDIA it was a Question why AMD not to try bigger memory interface then?.....
After all why NVIDIA is more powerful with its Quadro series for CAD works and etc. Yes they may have agreements with Adobe or Autodesk or anyone else. They bought Ageia, they being coordinated with software companies with something named CUDA... but in practice who (Pro person with enough money) wants to use FirePRO instead of Quadro for video editing? These are questions anyway.
Thank you again....
a b U Graphics card
March 24, 2011 1:02:41 PM

lol, you can't just take someones word well have fun
a b U Graphics card
March 24, 2011 1:03:19 PM

ct1615 said:
because its not an important factor. nvidia and AMD engineer their cards differently. sometimes they use the bus to slow a card down in order to meet a market segment, sometimes increasing the bus will show no improvement at all so there is no point.

the AMD has two gpu cores, each core has a bus width of 256. there is not a single core with 512 in the 6990.



agreed, the first post should have answered your question
Anonymous
a b U Graphics card
a b } Memory
March 27, 2011 6:53:45 AM

Best answer selected by bit0.
March 27, 2011 8:40:24 AM

Quote:
Hi,
Ati was the first to use GDDR5 on it's graphics cards I think,
but they still use 256 bit memory interface where NVIDIA Uses 320/384-bit memory interface at its high-end.
Please tell me why AMD seems to forget about this very important factor.
Another question is that when they say 512-bit memory interface (2x256 bit) for cards like 6990 is it right that we ADD two interfaces together? the efficiency of 2x256 lane is equal to 512?
Thanks.


Here's how it works in detail.

6990 is 2x6950 2GB built on one board (called PCB hereafter) with a bridge chip in between. In crossfire setups, or SLI, each GPU must map it's memory to the complementary GPU as well. This means that effectively the 6990 has 2GB VRAM at it's disposal as 1GB VRAM for each GPU is eaten up by the other to ensure consistency between frame buffers (VRAM). The cross-talk, or 'data sharing' between the GPUs adds some latency to the rendering of each frame.

The memory technologies used with each GPU don't affect this directly, although in theory faster memory speeds might reduce this cross-talk latency a bit.

Since the 6950 uses 256-bit GDDR5, it has the capability in some scenarios to read and write data from the frame buffer (VRAM) twice per cycle, expounded upon the DDR technology already in use in standard RAM and version of VRAM previous to GDDR5.

This means that in some scenarios, 256-GDDR5 is effectively a 512-bit DDR memory channel (keep in mind these bit bandwidth numbers are derived from using multiple memory chips at once in 32-bit or 64-bit configurations.)

So, after learning that, what you can see is that the 6950 uses 8 32-bit memory modules on individual channels that can send/receive data twice per clock most of the time.

To answer the question more thoroughly, does 256-bit GDDR5=512-bit DDR? Most of the time yes, but not all the time. GDDR5 does give a real performance boost, but it's not a 100% boost 100% of the time. 6990 is really two 6950 256-bit GDDR5 GPUs on the same PCB, that is all.

Hope this helps!

!