Official AMD Radeon 6000 Series Discussion Thread

[crazyeyesreaper]

the 6800 seem to offer better scaling in xfire then 5800s it might have something to do with the lower number of SPs there trying to keep data fed to as the easier it is for the cpu to push data to the better the cards should preform im guessing the overhead for xfire is reduceded depending on # of sp's so thus lower overhead means better scaling across the board

 

[CobraXP]

This might help some people to decide on 6870s. This is my fastest to date with cpu at 4.2GHz with turbo on

 

[pantherx12]

do you have 6800's? from all the reviews I've seen they perform as-good-as-on-par with their 5800 counterpart.

eh?

In the games that do scale it's typically above 85% the average actually being 92% scaling or something.

even including crappy scaling and benches it's still 70% average, which is better than anything ati has done before, cept when they done alternate frame rendering XD

 

[wahdangun]

the 6800 seem to offer better scaling in xfire then 5800s it might have something to do with the lower number of SPs there trying to keep data fed to as the easier it is for the cpu to push data to the better the cards should preform im guessing the overhead for xfire is reduceded depending on # of sp's so thus lower overhead means better scaling across the board

are you sure ? if that what happen then why HD 5770 didn't have amazing scaling like HD 68XX after all they even have FEWER SP than HD 68XX

 

[reverze]

I made a post in another thread, but I'm working on pushing my 6870 to the extreme..

Got this one off of NewEgg and have had amazing results:
http://www.newegg.com/Product/Product.aspx?Item=N82E16814150506

I'm thinking about going cold with it.. any suggestions ?

 

[20mmrain]

do you have 6800's? from all the reviews I've seen they perform as-good-as-on-par with their 5800 counterpart.

He might but I do. I can vouch for the fact that with the 10.10c hot fix drivers the cards do scale better then the 5800 series.
For example....2 5870's at stock speeds and with a i7 get about P27k overall in Vantage. My 6870's at stock speeds and a i5 650 @ 4.2 Ghz get P28k overall.

GPU scores with two 5870's is about 26k to 27k stock. With my two 6870's I score about 27k to 28k stock and that's not even with my CPU's highest overclock yet. If I switched to a i7 I believe I could be higher as well.

My point is that a 5870 is faster (Right now by a little) then a 6870 in single mode. But now with it's improved Xfire scaling.... a 6870 Xfire beats a 5870 Xfire.

At least in my experience. Yes I have owned both sets of cards too

 

[wolf]

cheers 20mmrain, my only point was he is quoted saying;

they perform so much better than the Hd 5800 series

and from the feedback they seem to perform just a tad better really, not like a night and day difference or anything.

 

[Benetanegia]

are you sure ? if that what happen then why HD 5770 didn't have amazing scaling like HD 68XX after all they even have FEWER SP than HD 68XX

They also have half the dispatch units. Both Cypress and Barts have dual dispatchers, Juniper does not and that of course makes a big difference. Cypress and Juniper both have 800 SP per dispatcher, while Barts has 560. Ever since Evergreen was released I always thought (me and many others) that they were bottlenecked by the dispatch and/or setup engine, the main reason being that the cut down versions (HD5x50) were as fast as their relatives when clocked at the same speed. IMO this was a direct consequence of the extra SPs that were enabled in RV770: according to a Anadtech article about the development of RV770, originally it would have 640 SP (2x RV670), but AMD realised they could put some more shaders in the space they got when getting rid of the ringbus. Despite RV770 having much more then 2x the number of shaders it was never really more than 2x faster, which even back in the day was a symptom of that small "bottleneck". Kick in DX11 and much more complex SPs, without any real change to the architecture and the dispatchers got even more stressed and became even more of a bottleneck. Barts fixed this by using a ratio lower than the originally planned ratio of 640, but it could have probably taken 640-720 just as well, but the (self-impossed) die area limitation prevented them for using more. In fact, AMD did have two Brts prototypes one with 32 ROP and 1120 SP and one with 16 ROP and 1280 SP. 1120 SP version won.

 

[the54thvoid]

Fudzilla (grains of salt etc) had this:

http://www.fudzilla.com/graphics/it...ecs?utm_source=twitterfeed&utm_medium=twitter

Don't know if it's news or not.

"Cayman packs 1536 stream processors, 32ROPs, 96 Texture units and 2GB of GDDR5 memory."

Also, UK prices of 5970 down to about £450. Clearing the way for a HD6970 with better performance and higher price?

 

[TheMailMan78]

They also have half the dispatch units. Both Cypress and Barts have dual dispatchers, Juniper does not and that of course makes a big difference. Cypress and Juniper both have 800 SP per dispatcher, while Barts has 560. Ever since Evergreen was released I always thought (me and many others) that they were bottlenecked by the dispatch and/or setup engine, the main reason being that the cut down versions (HD5x50) were as fast as their relatives when clocked at the same speed. IMO this was a direct consequence of the extra SPs that were enabled in RV770: according to a Anadtech article about the development of RV770, originally it would have 640 SP (2x RV670), but AMD realised they could put some more shaders in the space they got when getting rid of the ringbus. Despite RV770 having much more then 2x the number of shaders it was never really more than 2x faster, which even back in the day was a symptom of that small "bottleneck". Kick in DX11 and much more complex SPs, without any real change to the architecture and the dispatchers got even more stressed and became even more of a bottleneck. Barts fixed this by using a ratio lower than the originally planned ratio of 640, but it could have probably taken 640-720 just as well, but the (self-impossed) die area limitation prevented them for using more. In fact, AMD did have two Brts prototypes one with 32 ROP and 1120 SP and one with 16 ROP and 1280 SP. 1120 SP version won.

One word. Paragraphs.

 

[Benetanegia]

One word. Paragraphs.

Sorry, paragraphs are out of stock at this moment. Try again in a few days.

 

[crazyeyesreaper]

my point of reference is 6800 scale better in xfire on AMD rigs then 5800s do by good 15-20% from the looks thats a pretty big difference in terms of scaling which is why i say they scale way better on an i7 rig not as big a difference but put the duo of cards in an AMD rig 6800s seem to do better hands down

 

[mastrdrver]

If I'm not wrong, Nvidia bought a whole lot of wafers. Buy more wafers = bulk discount. Of course it is obvious that they don't make as much money per wafer as the competition due to the chips size.

Bulk discount only helps if GF104 yielded what Cypress did. It also depends on how you are paying: by working die or by the whole wafer.

It's not bigger, it's about the same size. 331 mm^2 vs 334mm^2.

Where's the source for the die size because as far as I know there has never been an official statement regarding size. Maybe I've just missed it or forgotten what I read (the later happens too often).

 

[Benetanegia]

Where's the source for the die size because as far as I know there has never been an official statement regarding size. Maybe I've just missed it or forgotten what I read (the later happens too often).

Most credible sources report 331/332 mm^2 die size on their reviews. For example Wizzard:

http://www.techpowerup.com/reviews/Colorful/iGame_GTX_460_1_GB/4.html

NVIDIA's GF104 graphics processor is made on a 40 nm process at TSMC Taiwan and is based on NVIDIA's Fermi architecture just like the more powerful GF100 on the GTX 480, for example. It uses approximately 1.95 billion transistors. Please note that the silvery metal surface you see is the heatspreader of the GPU. The actual GPU die is sitting under the heatspreader, and is roughly 332 mm² in size.

But also, Anandtech, Techreport, X-bit and countless of other respectable sites.

 

[wahdangun]

They also have half the dispatch units. Both Cypress and Barts have dual dispatchers, Juniper does not and that of course makes a big difference. Cypress and Juniper both have 800 SP per dispatcher, while Barts has 560. Ever since Evergreen was released I always thought (me and many others) that they were bottlenecked by the dispatch and/or setup engine, the main reason being that the cut down versions (HD5x50) were as fast as their relatives when clocked at the same speed. IMO this was a direct consequence of the extra SPs that were enabled in RV770: according to a Anadtech article about the development of RV770, originally it would have 640 SP (2x RV670), but AMD realised they could put some more shaders in the space they got when getting rid of the ringbus. Despite RV770 having much more then 2x the number of shaders it was never really more than 2x faster, which even back in the day was a symptom of that small "bottleneck". Kick in DX11 and much more complex SPs, without any real change to the architecture and the dispatchers got even more stressed and became even more of a bottleneck. Barts fixed this by using a ratio lower than the originally planned ratio of 640, but it could have probably taken 640-720 just as well, but the (self-impossed) die area limitation prevented them for using more. In fact, AMD did have two Brts prototypes one with 32 ROP and 1120 SP and one with 16 ROP and 1280 SP. 1120 SP version won.

but isn't HD 4870 have more than 2X performance than HD 3870 ?

 

[Duckula]

We managed to get some specs of the soon-to-launch, well at least on paper, AMD Radeon HD 6970 card.

What we can share with you is that Cayman packs 1536 stream processors, 32ROPs, 96 Texture units and 2GB of GDDR5 memory.

We, the hard working people of Fudzilla, also have learned that the card comes with two DVI ports, 2 mini Display ports and HDMI, which is much better than what Nvidia's GTX 580 can offer.

TDP looks to be slightly lower than GTX 580’s but in worst case it looks to go up to the sky, something that we’ve seen with GTX 480 and almost 300W TDP in Furmark. Stay tuned, more specs will follow as soon as we can confirm them.

http://www.fudzilla.com/graphics/item/20759-some-cayman-radeo-hd6970-specs

 

[Super XP]

my point of reference is 6800 scale better in xfire on AMD rigs then 5800s do by good 15-20% from the looks thats a pretty big difference in terms of scaling which is why i say they scale way better on an i7 rig not as big a difference but put the duo of cards in an AMD rig 6800s seem to do better hands down

Agreed!!

 

[X800]

Im really disapointed in 6870 (not preformance way) but its too loud its like the old 2900xt hairdryer.

 

[Zen_]

Im really disapointed in 6870 (not preformance way) but its too loud its like the old 2900xt hairdryer.

Have you tried tweaking the fan profile at all?

Part of the reason I decided on the 6850 over the 70 was the heatsink. The blower style heatsinks are nice for pushing air out the slot but they can get quite noisy.

 

[bear jesus]

Im really disapointed in 6870 (not preformance way) but its too loud its like the old 2900xt hairdryer.

Do you have a very hot case or something? i ask as my 6870's hardly make any noise while being stressed never mind when gaming or idle and although my use them in the open my ambient temps are usually between 22c and 30c.

 

[cadaveca]

I know my case is hot. I like to rub against it all the time...gives me the motivation to UPgrade.

 

[erocker]

Im really disapointed in 6870 (not preformance way) but its too loud its like the old 2900xt hairdryer.

That really shouldn't be. Is it possible that the fan itself has a bad bearing in it?

 

[the54thvoid]

http://www.xbitlabs.com/news/video/...tion_of_Dual_Chip_Flagship_Graphics_Card.html

6970 and 6950 still on track for Q4 2010 release. Antilles Q1 2011.

 

[yogurt_21]

http://www.xbitlabs.com/news/video/...tion_of_Dual_Chip_Flagship_Graphics_Card.html

6970 and 6950 still on track for Q4 2010 release. Antilles Q1 2011.

visual chart in there puts naming scheme in a better light. also shows that the 6970 isn't meant to beat the 5970 which is in line with what I was thinking

edit:

Considering the fact that Nvidia Corp. has just released its new flagship GeForce GTX 580, which offers up to 30% of performance increase over GeForce GTX 480 and generally offers a number of advantages compared to Radeon HD 5970, formally the highest-end consumer graphics card on the market, the delay of the Radeon HD 5990 is not a good news for AMD

well first they put 5990 and not 6990, then they said the 580 is up to 30% faster than the 480, abit misleading when W1z shows it averages 10% faster, and finally the 5970 is still faster than the 580, so an antilles delay will not affect amd at all.

 

[Benetanegia]

well first they put 5990 and not 6990, then they said the 580 is up to 30% faster than the 480, abit misleading when W1z shows it averages 10% faster, and finally the 5970 is still faster than the 580, so an antilles delay will not affect amd at all.

It is up to 30% faster. up to is not the same as average. On some games at certain settings it is easily 20-30% faster. The 30% claim comes from Vantage, I think, and that boost comes from the doubled FP16 capabilities on the texture units.

W1zzard's average numbers are not accurate to say how much faster a card can be, because many games, especially the ones using old engines (World of Warcraft, UT3, Supreme Commander) do not scale properly and that skewes the results for faster cards a bit. There's no any better metric on the net to say which is the average performance of a card than W1zzard's numbers, because tey do represent what the majority of people will get, but it's not a perfect one to prove/disprove "up to X% faster claims". I don't know if I have explained myself well.

Also even in Wizzard's average numbers it is 15% faster at 1920x1200 and 16% faster at 2560x1600 and even at "all resolutions" it is 11% faster. You have to divide the highest number by the lowest number, you can't just substract them because the chart is relative to only one card. Otherwise, HD5970 is 107% faster than the GTX480?