Looking for info on a 512-bit GDDR3 Ram at 1 ghz....
- Thread starter Josh378
- Start date
TOrangeMonkey
Newcomer
If it is, it's not for ATi, Huddy specificaly said it's not comming at GDC.Josh378 said:
Richard Huddy - ATI devrel (think he's devrel for Europe)
http://www.hynix.com/datasheet/eng/...14&RAM_NAME=GDDR3 SDRAM&SUB_RAM=256Mb
hynix also at 800mhz either 256Mbit or 512Mbit chips with 1600Mbit/s per pin
hynix also at 800mhz either 256Mbit or 512Mbit chips with 1600Mbit/s per pin
It's 32bit per GDDR3 memory chip. That's why you need at least 8 of them for a 256bit interface. You could use more on a wider interface, but that means larger PCB, more traces (= more layers), more pins on the GPU. In short: higher cost. But it's doable.Josh378 said:
Joe DeFuria
Legend
Where are 1.6 Ghz GDDR-3 parts expected to be on the market? :?
Josh...there may be some misunderstanding between you, your contact, and the industry on terminology or conventions when talking about ram speed, density, etc. (It may not be that you were intentionally lied to.)
Josh...there may be some misunderstanding between you, your contact, and the industry on terminology or conventions when talking about ram speed, density, etc. (It may not be that you were intentionally lied to.)
Joe DeFuria said:
It's not my strength of expertise...but I'm getting there in terms of education...
I was told that GDDR3 right now has a 512-bit at 1 ghz ready right now...which is kinda strange that they haven't released to the public yet of a 256-bit GDDR3 at 1066GHZ yet...
http://www.xbitlabs.com/news/memory/display/20050218041035.html
Samsung Electronics, the world’s largest maker of dynamic random access memory, said Thursday it had made the world’s first 512Mb memory chip which complies to next-generation GDDR3 standard and can operate at the speed on 1066MHz. The company says the new memory type may reach commercial production as early as in 2006.
I was kinda confused...so i came here to ask you guys on your opinion since we have professionals here who know soo much more than I do...
-Josh378
Joe DeFuria
Legend
Josh378 said:
It's certainly understandable how you can be confused...turning on my marketing translator: the above doesn't even mean they have created a product sample with the mentioned specs. It means they have some new chip design in a lab with 512Mb density...and that they anticipate this memory tpye top speed to be 1066 Mhz.
I wouldn't even read into that as 1066 Mhz. will be available as early as 2006.
(Oh, and 1066 Mhz is not 1.6 Ghz...it's 1.06. )Kinda similar to Intel showing "high speed GPUs" that do little more than "boot into Windows" using exotic cooling methods.
Joe DeFuria
Legend
Well, I would guess that for graphics cards, the fastest we'll see in shipping products is what is sampling today: 800 Mhz. This would give 51 GByte/sec bandwidth on a board level, assuming 256 Bit wide bus for the graphics chips.
There may be faster modules (1 Ghz? or GDDR4?) smpling or available in smallish quantities, but I doubt we'd see them in products this year.
There may be faster modules (1 Ghz? or GDDR4?) smpling or available in smallish quantities, but I doubt we'd see them in products this year.
one more thing, can someone explain this to me please:
Each of the Realizm 800’s visual processors interfaces with a 256-bit memory bus populated by 256MB of GDDR-3 memory, forming what 3DLabs likes to call a 512-bit bus.
http://reselleradvocate.com/public/ram/issues/ram36/ram36_reloaded01.html
So this means that were getting a TRUE 512-Bit GDDR3 or what? This will be my last question...I apologies if I'm wasting your time guys...
-Josh378[/i]
Each of the Realizm 800’s visual processors interfaces with a 256-bit memory bus populated by 256MB of GDDR-3 memory, forming what 3DLabs likes to call a 512-bit bus.
http://reselleradvocate.com/public/ram/issues/ram36/ram36_reloaded01.html
So this means that were getting a TRUE 512-Bit GDDR3 or what? This will be my last question...I apologies if I'm wasting your time guys...
-Josh378[/i]
Josh378 said:
I think you are confusing memory density (Mbit) with memory interface width (bits). 3DLabs is saying that this graphics processor, which is really two of them, each having a 256-bit memory interface, combine to form what they call a 512-bit memory interface. So this is not a true 512-bit interface as we know it. This would be something like Nvidia claiming a 512-bit memory interface when using SLI. It's not really true in practice. However, it is true that the number of bits used (wide) in all (in this case 2) memory interfaces on the device amount to 512. (* See EDIT)
Just to iron this out for you, when DRAM manufacturers speak about a 512-Mbit memory chip they are talking about how many bits it has. So divide this by 8 (8 bits/byte) and you get Mbytes.
512Mbit/8 = 32MB
In a 256-bit memory crossbar, with each 'channel' or 'lane' constituting 32 bits, you need 8 (8*32 = 256). So multiply 8 * 32MB and you see the maximum configuration with this 512Mbit chip turns out to be 512 MB.
So GDDR3 chips are 32 bits wide and you need 8 of them to fully populate a 256-bit memory interface. For a 512-bit interface you would either need 16 of these DRAM chips or some form that is wider, like 64-bit wide modules (you could imagine each chip being dual channel).
EDIT:
It should probably be considered that the modern 3D chip is sectioned up to some extent in groups of 4 pixel pipelines called quads. So, one could imagine that the memory interface could also be designed to service these quads (where Radeon 9700 has 8 and Geforce 6800GT/Ultra have 16). So, the fact that the chip is two 'units' 'glued' together may not be important when speaking about the width of the bus. I was a bit hasty and rough comparing this situation to Nvidia SLI, because SLI has lots of other issues that may not be present on the Realizm. So, it's probably best to concede that this is a 512-bit interface unless there is some information particular to this design that would reveal that it never really acts like a unified memory interface with 512-bits would. I'm considering the possibility that a non-unified memory interface may be just as good, or perhaps even better, than a unified interface when it comes to these moder, highly scalable architectures.
Similar threads
