Monday, November 21st 2011

Noises About Radeon HD 7900 Series with XDR2 Memory Grow

As early as in September, we heard reports of AMD toying with Rambus XDR2 memory on its next generation of high-performance GPUs. Apart from our own community's response, that news met with a wall of skepticism as it was deficient in plausibility. New reports from Chinese websites have raised the topic again with fresh rumors that AMD will attempt to implement XDR2 on some of its next-generation ultra-high end products after all. XDR2, according to Rambus, can transport twice the amount of data per clock as GDDR5.

Apparently AMD and Rambus have had much more cordial relations with each other, than other companies the latter engaged in patent disputes with. In 2006, AMD settled outstanding disputes with Rambus by willing to pay licensing costs for certain technologies claimed by Rambus, turning a leaf in the relations between the two. What Chinese sources are suggesting now, is that AMD will design its high-end GPU (codename: "Tahiti") in a way that will let it support both GDDR5 and XDR2. Certain higher-end SKUs based on Tahiti will use XDR2, while the slightly more cost-effective SKUs will use GDDR5.

In related news, other sources told TechPowerUp that AMD could adopt a "top-to-bottom" strategy with the high-end portion of its next-generation of products. This means that AMD could launch the dual-GPU "New Zealand" graphics card first, followed by single-GPU SKUs.Source: Mydrivers
Add your own comment

64 Comments on Noises About Radeon HD 7900 Series with XDR2 Memory Grow

#1
EastCoasthandle
John Doe said:
And it doesn't, there's nothing limiting the card from using XDR on PCI-E 2.0 x16.



Really? I'm clearly ignorant eh? Way to go buddy. Yes, they ARE kinds of SDRAM. But they DIFFER between each. Full stop.
LOL, I'm talking about a 7970 with XDR2 vs 7970 using GDDR5 in current vs next gen PCIe. I should have made "it" more clear but that's what I was referring to.
Posted on Reply
#2
Unregistered
cadaveca said:
IOMMU in 7-series GPU. That is all.

Pretty sure ECH was thinking about this when mentioning PCI-E. I am sure he was wondering is PCIe bandwidth was enough to take full advantage of the bandwidth offered by XDR2.

Not all users use VGAs for gaming..GPGPU has it's purposes too.
EastCoasthandle said:
LOL, I'm talking about a 7970 with XDR2 vs 7970 using GDDR5 in current vs next gen PCIe. I should have made "it" more clear but that's what I was referring to.
PCI-E has sufficient bandwidth to make use of memory. Do you have an idea on how much bandwidth PCI-E x16 2.0 can provide? How much of it is used by memory? A miniscule amount.
Posted on Edit | Reply
#3
largon
btarunr said:
You are currently running DDR3-SDRAM.
Right, instead of SDR-SDRAM (Single Data Rate Synchronous Dynamic Random Access Memory).
John Doe obviously referred to data rate with "SD" (SDR-SDRAM), not the basic type of the DRAM.

You're being a pedant for no reason.
Posted on Reply
#4
EastCoasthandle
John Doe said:
PCI-E has sufficient bandwidth to make use of memory. Do you have an idea on how much bandwidth PCI-E x16 2.0 can provide? How much of it is used by memory? A miniscule amount.
I see, LOL you must be having a bad day arguing with people here. Let me put it this way as I really find your reply silly. I want to see the benchmark results and find your reply pointless.
Posted on Reply
#5
Unregistered
EastCoasthandle said:
I see, LOL you must be having a bad day arguing with people here. Let me put it this way as I really find your reply silly. I want to see the benchmark results and find your reply pointless.
Does DDR to GDDR3 make difference from AGP to PCI-E x16? It doesn't. Memory as a whole doesn't matter to a significant amount. It doesn't make enough difference to be noticeable from one type to another, let alone from one interface to the other. What matters is the GPU itself; it's core. Not the memory, which is why your arguement is pointless.
Posted on Edit | Reply
#6
cadaveca
My name is Dave
Memory controller interfaces with system ram, via DMA, over PCIe bus.

XDR2 has it's own specialized controller, which may affect how system memory is accessed. PCIe 3.0 offers less overhead, which may make transfers faster over PCIe 2.0, if done right.

Add in the IOMMU, for additional system ram accesses, and yes, the interface used might play a role.


I'm not sure why you think it's a moot point. Nothing wrong with answering a question, now is there?
Posted on Reply
#7
Unregistered
cadaveca said:
Memory controller interfaces with system ram, via DMA, over PCIe bus.
System RAM is sufficiently fast. Like I said, nothing is limiting XDR2. For a fact, nothing is limiting PCI-E 2.0 at all. That's why 3.0 still isn't set up yet.

cadaveca said:
XDR2 has it's own specialized controller, which may affect how system memory is accessed. PCIe 3.0 offers less overhead, which may make transfers faster over PCIe 2.0, if done right.

Add in the IOMMU, for additional system ram accesses, and yes, the interface used might play a role.
You're talking as if mem performance matters over what we currently have. Not to a significant amount, and 2.0 already has low overhead.
Posted on Edit | Reply
#8
EastCoasthandle
John Doe said:
Does DDR to GDDR3 make difference from AGP to PCI-E x16? It doesn't. Memory as a whole doesn't matter to a significant amount. It doesn't make enough difference to be noticeable from one type to another, let alone from one interface to the other. What matters is the GPU itself; it's core. Not the memory, which is why your arguement is pointless.
Your post is pointless. You can't reference older ddr or gddr3 to AMD's next gen video cards using current hardware. Which is why I said the tests should be done. Get mad all you want but your entire reasoning, referencing old hardware is the very reason why I made the suggestion. The only way to know what the results are is if it's tested once it's made available. :laugh:
Posted on Reply
#9
cadaveca
My name is Dave
John Doe said:
For a fact, nothing is limiting PCI-E 2.0 at all.
System ram efficiency affects GPU performance. Just fire up a CodeMasters title to see what I mean. I use F1 2010 in my board reviews specifically because it's so sensitive to system ram performance.

John Doe said:
You're talking as if mem performance matters over what we currently have
The addition of an IOMMU into the GPU itself does make the memory subsystem a bit more important for GPGPU. Gaming...time will tell.
Posted on Reply
#10
Unregistered
EastCoasthandle said:
Your post is pointless. You can't reference older ddr or gddr3 to AMD's next gen video cards using current hardware. Which is why I said the tests should be done. Get mad all you want but your entire reasoning, referencing old hardware is the very reason why I made the suggestion. The only way to know what the results are is if it's tested once it's made available. :laugh:
What's done will be done. The card isn't out yet, so the closest comparison I can do would be with older hardware.

cadaveca said:
System ram efficiency affects GPU performance. Just fire up a CodeMasters title to see what I mean.
Bullshit.

http://www.anandtech.com/show/2792
Posted on Edit | Reply
#11
cadaveca
My name is Dave
John Doe said:

Bullshit.

http://www.anandtech.com/show/2792
Reading comprehension in important. I said Bandwidth efficiency, not speed. There is very much a huge difference between the two.




Slower CPU, but faster memory, wins. Note that both are using identical ram speeds(1333 MHz 9-9-9-24). The APU platform is far more efficient at memory control than Thuban's.

I make a point of using benchmarks anyone can replicate.
Posted on Reply
#12
Unregistered
cadaveca said:
Reading comprehension in important. I said Bandwidth efficiency, not speed. There is very much a huge difference between the two.
Neither changes anything in this case since both PCI-E 2.0 and DDR3 already have a fast balance.

cadaveca said:
Slower CPU, but faster memory, wins. Note that both are using identical ram speeds(1333 MHz 9-9-9-24). The APU platform is far more efficient at memory control than Thuban's.

I make a point of using benchmarks anyone can replicate.
That doesn't have anything to do. It's comparing a different CPU memory controller to the other, not lane controller or faster memory to slower.
Posted on Edit | Reply
#13
cadaveca
My name is Dave
John Doe said:
That doesn't have anything to do. It's comparing a different CPU memory controller to the other, not lane controller or faster memory to slower.
That's exactly my point. XDR2 has it's own controller, which will use the PCIe bus to communicate with system ram. Like how the APU controller makes better use of the bandwidth available, likewise, XDR2's controller may as well. This may place greater load on the PCIe bus.

This is doublied by the addition of an IOMMU, which is basically just yet another memory controller, but one specific to System Memory Access, rather than both accessing system memory, and onboard caches. Perhaps the use of the IOMMU is needed with the XDR2 contorller, or perhaps not...time will tell.


It's something to be aware of, and to look at, when and if this new technology is used. This also bolster's ECH's interest in if PCIe matters..he never said it would matter...he was curious how it will play out. Could amount ot nothing, of course, but it's far to early to tell.
Posted on Reply
#14
chron
John Doe said:
And it doesn't, there's nothing limiting the card from using XDR on PCI-E 2.0 x16.



Really? I'm clearly ignorant eh? Way to go buddy. Yes, they ARE kinds of SDRAM. But they DIFFER between each. Full stop.
He's just saying you were ignorant about one small fact, not in general.
Posted on Reply
#15
Unregistered
chron said:
He's just saying you were ignorant about one small fact, not in general.
No, I wasn't. I referred to "SD" and "SDR" RAM. Not as DDR RAM. It's only uninformed if you don't know what you're reading.
Posted on Edit | Reply
#16
OneCool
chodaboy19 said:
It's only a matter of time before RAMBUST sues AMD too. LOL
They did
Posted on Reply
#17
theubersmurf
The prices Rambus charges puts their products out of the consumer range. This isn't happening, I'm confident. This is just scuttlebutt.
Posted on Reply
#18
mastrdrver
There is no manufacturer of memory currently even making XDR2. So how are you suppose to validate a memory controller without any memory chips? How are you suppose to make cards without any memory to put on them so you can validate the card?

I wish there was a black hole on the internet where we could put this XDR2 rumor so it would never return. :banghead:
Posted on Reply
#19
chron
mastrdrver said:
There is no manufacturer of memory currently even making XDR2. So how are you suppose to validate a memory controller without any memory chips? How are you suppose to make cards without any memory to put on them so you can validate the card?

I wish there was a black hole on the internet where we could put this XDR2 rumor so it would never return. :banghead:
The industry is probably at a point where they can see where things are going with a certain kind of technology, even when it's at it's infant stage.
Posted on Reply
#20
Casecutter
chodaboy19 said:
It's only a matter of time before RAMBUST sues AMD too. LOL
Yep they had litigation, but we weren’t actually made privy to what to an exact settlement. Could have RAMBUS and AMD worked out an agreement for use of technology? RAMBUS honestly at some point needs to find someone, they aren’t friendly with Nvidia so AMD is a likely partner.

theubersmurf said:
The prices Rambus charges puts their products out of the consumer range. This isn't happening, I'm confident. This is just scuttlebutt.
Well that was back a long time ago, and then it really never made it to market so hard to say that position is valid today with the right partner. While it will be the premium offering on the highest end offer card. If it works... there will be takers

mastrdrver said:
There is no manufacturer of memory currently even making XDR2. So how are you suppose to validate a memory controller without any memory chips? How are you suppose to make cards without any memory to put on them so you can validate the card?
So what’s Rod Dhat Solution Marketing Manager for RAMBUS using? How do you know who is or who’s not making chips? Then why did AMD seemingly start branding Radeon memory that might have been Hyundai (Hynix)? Your entire dither presents the dumbest argument. You're saying a company like AMD can’t find somewhere/someone to make R&D samples and production? ...wow is your thinking that pathetic! :wtf:
Posted on Reply
#21
MarcusTaz
Jeeze will you all stop your bitching... :roll:

IF this thing has XDR2 I want it !!! :laugh:
Posted on Reply
#22
Relayer
Because of the extra bandwidth not only would 256 mem bus be still usable on next generation high end, but 128 could be used on GPU's as powerful as the current 6970 making the GPU's smaller and cheaper. Since AMD sells GPU's (at least after the initial release of reference cards) not complete cards or the RAM chips, it would make the product they sell cheaper to produce. They've already paid the licensing, which I read runs out this year, I see no reason for AMD not to do it. Providing of course it's readily available and the engineering isn't so hard as to hold back production.

Add to that AMD seems to like to be ahead of the curve vs. nVidia with new tech.
Posted on Reply
#23
Unregistered
Relayer said:
Add to that AMD seems to like to be ahead of the curve vs. nVidia with new tech.
Yeah, that's what you do when your GPU isn't as powerful. Fanboyism is stupid ain't it?
#24
de.das.dude
Pro Indian Modder
John Doe said:
System RAM is sufficiently fast. Like I said, nothing is limiting XDR2. For a fact, nothing is limiting PCI-E 2.0 at all. That's why 3.0 still isn't set up yet.



You're talking as if mem performance matters over what we currently have. Not to a significant amount, and 2.0 already has low overhead.
so you would prefer a 6990 with 4 gigs of cheap ass DDR3 rather than GDDR5.
Posted on Reply
#25
eidairaman1
The Exiled Airman
Reason GDDR5 is used is because it can get more data thruput over a 256bit bus and still be cheaper than GDDR3 or DDR3 over a 512bit bus.
Posted on Reply
Add your own comment