HD6900 series to ship with 2GBs of Ram

[dangerman1337]

Its possible but the logic involved in getting there may be a bit too narrow to assign it much probability.

Consider the loss in IPC for Evergreen over its predecessor. We never really came to a satisfactory resolution why an HD5870 under-performs an identically clocked xfire HD4870 setup.

So we know there is opportunity for AMD's engineers to "fix" whatever architectural bottleneck was introduced with evergreen that led to that particular decrease in IPC with the Evergreen architecture.

In addition there is the non-zero possibility that they further improved on Cayman's architecture (IPC) above and beyond what they once had with R600.

Obviously there are limitations to the analogy but consider PIII -> P4 -> C2D for Intel...we know IPC can shift drastically with architecture changes so it is not unreasonable to expect cayman to have the potential to boost IPC.

And what we don't know about the 40% is are the clockspeed deltas factored into that number? 40% IPC delta seems unreasonable, but a combination of IPC + clockspeed bumps could bring that kind of end-performance improvement.

I agree it is at the furthest reaches of plausibility...but nothing about the numbers render them entirely implausible. If the claim was 100% improvement or some such then yeah we could wholly write that rumor off as BS.

Didn't rumours put northern islands being a hybrid of R600/cyrpess and Islands arch?

 

[dguy6789]

Its possible but the logic involved in getting there may be a bit too narrow to assign it much probability.

Consider the loss in IPC for Evergreen over its predecessor. We never really came to a satisfactory resolution why an HD5870 under-performs an identically clocked xfire HD4870 setup.

So we know there is opportunity for AMD's engineers to "fix" whatever architectural bottleneck was introduced with evergreen that led to that particular decrease in IPC with the Evergreen architecture.

In addition there is the non-zero possibility that they further improved on Cayman's architecture (IPC) above and beyond what they once had with R600.

Obviously there are limitations to the analogy but consider PIII -> P4 -> C2D for Intel...we know IPC can shift drastically with architecture changes so it is not unreasonable to expect cayman to have the potential to boost IPC.

And what we don't know about the 40% is are the clockspeed deltas factored into that number? 40% IPC delta seems unreasonable, but a combination of IPC + clockspeed bumps could bring that kind of end-performance improvement.

I agree it is at the furthest reaches of plausibility...but nothing about the numbers render them entirely implausible. If the claim was 100% improvement or some such then yeah we could wholly write that rumor off as BS.

Just want to point out this interesting chart that caught my eye back when the 6870 article was published.

http://images.anandtech.com/graphs/graph3987/33253.png

First time ever that the 5870 had well over a 100% performance advantage over the 4870? Any speculation for why that is?

 

[StrangerGuy]

"HD6900 series will have 2GBs of Ram "

Fixed for ya'. (seriously who doesn't expect that...)

 

[Arkadrel]

Hold on there. Are they saying that Cayman is essentialy going to be the same size as Cypress, but 40% faster on the same proccess?

The Cypress is 334mm^2 as far as I know.
The Barts is 255mm^2. (like +31% smaller, but only 5-10% off in performance of the 5870).

a 360mm^2 is ~42% bigger than a Barts (6870).

Now lets say they found some small bottleneck, fixed that problem, and added architectural changes that save die space + give boost in performance (like say 4D shaders). And if all else fails clock it kinda high to reach expected results.

I could see Amd doing a 360mm^2 card that has more than +60% performance than a 6870.

The 580 is in the same boat, it ll be 20% ish faster than a 480 (rumor sites)... which will put it at about 40-50% faster than a 5870.


Rumors ("supposed" benchmark leaks from chinese sites like chiphell) put the 580 and the 6970 very close to one another.

The question is... has Nvidia made a smaller chip than the 480's 529mm^2 ?

Or will it use a 529mm^2 to combat a 360mm^2 chip (that probably costs 47%-50% less to make (due to size differnce = chips pr waffer) and because of bigger = less yields).

 

[dookulooku]

The Cypress is 334mm^2 as far as I know.
The Barts is 255mm^2. (like +31% smaller, but only 5-10% off in performance of the 5870).

There's nothing really special about Barts. It's just a more efficient configuration (shader/texture/rop/bandwidth) of the Evergreen architecture, along with improvements to achieve higher clocks at lower power and reduce size. AMD was off in their estimates for Cypress and put in too many SIMDs. HD5850 was less than 10% slower than 5870 at the same clocks, despite having 20% fewer SIMDs. 6870 has 30% fewer SIMDs, but higher clock speeds, which explains why it's performance is so close to 5870's.

a 360mm^2 is ~42% bigger than a Barts (6870).

Now lets say they found some small bottleneck, fixed that problem, and added architectural changes that save die space + give boost in performance (like say 4D shaders). And if all else fails clock it kinda high to reach expected results.

I could see Amd doing a 360mm^2 card that has more than +60% performance than a 6870.

I don't think we'll see +60% performance across the board. It's not going to be possible without large increases in ROP count and memory bandwidth. There's just not much to be gained by adding more shaders or increasing their efficiency.

I could be wrong, but it's most likely we'll see big improvements in tessellation and compute performance, areas where AMD is behind Nvidia.

 

[Arkadrel]

well...

6870 has 134.4 Gb/s memory bandwidth.
256bit bus width + GDDR5 @6 Ghz (rumored) = 192.0 Gb/s memory bandwidth.

So it ll likely have 43% more memory bandwidth than a 6870.

Rumors say their still at the same # of RoPs..... anywoo... chiphell are usually right about "leaks"... if their to be believed it ll be a really close race between thet 580 and the 6970.

 

[taltamir]

yeah, amd can get the same bandwidth with 256 bit that nvidia gets with 384 due to their better memory controller.

I thought it was because they used more expensive ram.

 

[Arkadrel]

I think overall the saved die space is cheaper than maybe a little more on the ram, if thats even the case.

Point is.. its more economical to use a smaller bit bus + fast ram if thats enough memory bandwidth (from what Ive heard ppl say on Semi accurate).

Which is why most assume that the reason nvidia arnt doing it is because they are haveing trouble makeing a smaller bit bus that can handle those speeds.

If you can make cheaper cards without performance loss why not do it? only sane reasoning = its tricky and giveing them trouble.

 

[GaiaHunter]

I thought it was because they used more expensive ram.

No.

http://www.xbitlabs.com/articles/video/display/geforce-gtx-480_2.html

The specifications suggest that these memory chips have an access time of 0.4 nanoseconds and a rated frequency of 5000 MHz. However, the memory frequency of the GeForce GTX 480 is only 3696 MHz, which leaves ample room for overclocking. To save power and dissipate less heat the card drops its memory frequency to 270 MHz in 2D applications. The graphics memory bus is 384 bits wide, providing an impressive peak bandwidth of 177.4 GBps.

 

[taltamir]

No.

http://www.xbitlabs.com/articles/video/display/geforce-gtx-480_2.html

The specifications suggest that these memory chips have an access time of 0.4 nanoseconds and a rated frequency of 5000 MHz. However, the memory frequency of the GeForce GTX 480 is only 3696 MHz, which leaves ample room for overclocking. To save power and dissipate less heat the card drops its memory frequency to 270 MHz in 2D applications. The graphics memory bus is 384 bits wide, providing an impressive peak bandwidth of 177.4 GBps.

Ok, so their chips are rated for much greater speeds than they are using them for... but this seems to indicate that this is a power saving decision, not a matter of superior controller

 

[blastingcap]

Ok, so their chips are rated for much greater speeds than they are using them for... but this seems to indicate that this is a power saving decision, not a matter of superior controller

NV struggled with their memory controller and this has been well-known since at LEAST March 2010: http://www.anandtech.com/show/2977/...-gtx-470-6-months-late-was-it-worth-the-wait-

"One thing that caught our eye with all of this was that NVIDIA’s memory clocks are lower than we had been initially expecting. GDDR5 is readily available up to 5GHz while NVIDIA doesn’t go any higher than 3.7GHz; in fact between the smaller memory bus than the GTX 285 and the lower than expected memory clocks, the GTX 400 series doesn’t have all that much more bandwidth than the GTX 285 did. As it stands the GTX 480 only has 11% more memory bandwidth than the GTX 285, while the GTX 470 has 15% less than the GTX 285.

Given the 384-bit bus, we initially assumed NVIDIA was running in to even greater memory bus issues than AMD ran in to for the 5000 series, but as it turns out that’s not the case. When we asked NVIDIA about working with GDDR5, they told us that their biggest limitation wasn’t the bus like AMD but rather deficiencies in their own I/O controller, which in turn caused them to miss their targeted memory speeds. Unlike AMD who has been using GDDR5 for nearly 2 years, NVIDIA is still relatively new at using GDDR5 (their first product was the GT 240 late last year), so we can’t say we’re completely surprised here. If nothing else, this gives NVIDIA ample room to grow in the future if they can get a 384-bit memory bus up to the same speeds as AMD has gotten their 256-bit bus."

P.S. AMD practically co-developed GDDR5, so of course they weren't going to struggle with it: http://www.anandtech.com/show/2679/8

 

[taltamir]

thank you, that clarifies the issue, I admit I was wrong.

 

[blastingcap]

thank you, that clarifies the issue, I admit I was wrong.

That's okay, not everyone camps out at tech sites and hits refresh like a rat on crack. Not that I'm... admitting... anything.... :whiste:

 

[GaiaHunter]

Ok, so their chips are rated for much greater speeds than they are using them for... but this seems to indicate that this is a power saving decision, not a matter of superior controller

Weren't you saying NVIDIA used cheaper ram?

I answered. They are using the same ram.

 

[taltamir]

That's okay, not everyone camps out at tech sites and hits refresh like a rat on crack. Not that I'm... admitting... anything.... :whiste:

I actually do that, but I was still wrong , must have missed that one issue, or maybe I even saw it and just forgot.

Weren't you saying NVIDIA used cheaper ram?

I answered. They are using the same ram.

and I said "ok", which means I accept what you said and that I was wrong in asserting that it is due to cheaper ram.
I followed with a "but", where I posited that, based on the info you posted, it might be a decision to reduce power consumption and temp rather then a problem with their controller.
blastingcap shot down that theory, so I admitted I was wrong and that nvidia indeed has a problem with their ram controller.

 

[GaiaHunter]

and I said "ok", which means I accept what you said and that I was wrong in asserting that it is due to cheaper ram.
I followed with a "but", where I posited that, based on the info you posted, it might be a decision to reduce power consumption and temp rather then a problem with their controller.
blastingcap shot down that theory, so I admitted I was wrong and that nvidia indeed has a problem with their ram controller.

I'll have to quote blastingcap.

not everyone camps out at tech sites and hits refresh like a rat on crack

That means I hadn't seen blastingcap explanation to you when I hit post.

Although I don't see where in the information I posted they say it is an energy saving feature, so I was a bit confused with your reply.

 

[Idontcare]

Didn't rumours put northern islands being a hybrid of R600/cyrpess and Islands arch?

I don't know who this "rumours" fella is, but if I ever catch him posting around here I'm a gonna slap his ass with a banhammer so big it'll make his grandchildren cry out in the name of jebus. :twisted:

Ouch! gah dam that smarts, even ma wee granchillen arr a cryen out for murcee from the wee baby jebus!

That'll learn ya! Now off with ye, back to S/A or wherever ye spawn from ya dastardly demonz!

Just want to point out this interesting chart that caught my eye back when the 6870 article was published.

http://images.anandtech.com/graphs/graph3987/33253.png

First time ever that the 5870 had well over a 100% performance advantage over the 4870? Any speculation for why that is?

NFC

 

[T2k]

You gotta love AMD's rumor spinners...

But hey, it works very well so far - a beautifully simple solution against leaks and industrial espionage.

 

[T2k]

NV struggled with their memory controller and this has been well-known since at LEAST March 2010: http://www.anandtech.com/show/2977/...-gtx-470-6-months-late-was-it-worth-the-wait-

"One thing that caught our eye with all of this was that NVIDIA’s memory clocks are lower than we had been initially expecting. GDDR5 is readily available up to 5GHz while NVIDIA doesn’t go any higher than 3.7GHz; in fact between the smaller memory bus than the GTX 285 and the lower than expected memory clocks, the GTX 400 series doesn’t have all that much more bandwidth than the GTX 285 did. As it stands the GTX 480 only has 11% more memory bandwidth than the GTX 285, while the GTX 470 has 15% less than the GTX 285.

Given the 384-bit bus, we initially assumed NVIDIA was running in to even greater memory bus issues than AMD ran in to for the 5000 series, but as it turns out that’s not the case. When we asked NVIDIA about working with GDDR5, they told us that their biggest limitation wasn’t the bus like AMD but rather deficiencies in their own I/O controller, which in turn caused them to miss their targeted memory speeds. Unlike AMD who has been using GDDR5 for nearly 2 years, NVIDIA is still relatively new at using GDDR5 (their first product was the GT 240 late last year), so we can’t say we’re completely surprised here. If nothing else, this gives NVIDIA ample room to grow in the future if they can get a 384-bit memory bus up to the same speeds as AMD has gotten their 256-bit bus."

P.S. AMD practically co-developed GDDR5, so of course they weren't going to struggle with it: http://www.anandtech.com/show/2679/8

Classic Nvidia-story, it has the same disastrous mis-management (by The Great Leader JHH) fingerprints all over it - ego and stupidity always trumps logic and openness there.

 

[RussianSensation]

"HD6900 series will have 2GBs of Ram "

Fixed for ya'. (seriously who doesn't expect that...)

Actually until recently there was no confirmation that HD6900 would ship with 2GBs of ram. Most speculation supported the notion that it might have 2GBs of ram, but 1GB was not discounted due to 256-bit memory bandwidth.

 

[cusideabelincoln]

Just want to point out this interesting chart that caught my eye back when the 6870 article was published.

http://images.anandtech.com/graphs/graph3987/33253.png

First time ever that the 5870 had well over a 100% performance advantage over the 4870? Any speculation for why that is?

5000 series probably has driver optimizations the 4000 series does not. Maybe SCII just really likes the slightly tweaked architecture of the 5000 series. Most likely a combination of things.

Also note that the 5770 is well outperforming the 4870, too.

 

[RussianSensation]

First time ever that the 5870 had well over a 100% performance advantage over the 4870? Any speculation for why that is?

Probably the ability of HD58xx series to run some parts of code much faster than HD4870. There are other games where this happens:

Mafia II

Poor HD4870 is reaching the end of the line. Epic card.

 

[dguy6789]

That's very cool. Perhaps the Cypress architecture was more future looking than once thought.

 

[RussianSensation]

That's very cool. Perhaps the Cypress architecture was more future looking than once thought.

More modern games also continue to become more demanding (esp. shader intensive where HD5870 is a power house). Therefore, the differences in performance between generations tend to grow since initial product launch compared to older generations.

Remember how in older games 8600GTS sucked and would be beaten by 7950GT/7900GTX? If you compare more modern games, 8600GTS > 7900 series.

 

[Skurge]

I don't know who this "rumours" fella is, but if I ever catch him posting around here I'm a gonna slap his ass with a banhammer so big it'll make his grandchildren cry out in the name of jebus. :twisted:



That'll learn ya! Now off with ye, back to S/A or wherever ye spawn from ya dastardly demonz!



NFC

Uhm, ok?

I think you need some rest there IDC.