6770 & 6750 Specs leaked

[LordEC911]

faster than rv840, but not as fast as rv870...

either memory chips that clock slightly lower are cheaper, or they are doing it on purpose... maybe it would be too close to rv970 otherwise... or maybe they are reserving some space for a refresh in case 28nm has problems like 40nm?

Barts has 5Gbps chips, just because you underclock doesn't mean they are cheaper. They are underclocked more than likely to save TDP since clocking them higher doesn't offer any substantial gains in performance, or at least not as much as core clock.

Edit- Also remember, AMDs word for this release is efficiency.

[Olivon]

Edit- Also remember, AMDs word for this release is efficiency.

Efficiency is one thing LordEC but performance is another thing. Barts needs to kick GTX 460 in the nuts

[LordEC911]

Efficiency is one thing LordEC but performance is another thing. Barts needs to kick GTX 460 in the nuts

It should.
Was just reminding people that raw numbers don't mean anything if you aren't using them efficiently.

[Olivon]

It should.
Was just reminding people that raw numbers don't mean anything if you aren't using them efficiently.

We are agree. If you look at size mm^2/performance, AMD did an excellent job with HD5K, especially regarding Fermi derivatives

edit : DX11, Tesselation need some improvements though

[Manicdan]

the gap between the 6870 and whatever 69x0 is the lowest they go with, might have been to small with full speed memory.

i just hope it overclocks VERY well.

[LightSpeed]

i just hope it overclocks VERY well.

Yeah, Barts supposedly does overclock VERY well

[antiacid]

so where's the 6970 leaks ?

[LordEC911]

so where's the 6970 leaks ?

Give it about a month.

[dengyong]

I have 7 figures in my bank and they're not my account number.

Are they all left of the decimal and on the same line ? j/k I'm glad you're doing well.

[Johnny87au]

SO the barts and XT arent better then the 5870? Anyone have any idea where their releasing the HD68XXX series??

[LordEC911]

SO the barts and XT arent better then the 5870? Anyone have any idea where their releasing the HD68XXX series??

Should be, at a minimum, very close in performance. New rumor is that Barts is actually performing like an early rumor, around GTX480 levels.

From the info I have heard, Barts IS the 68x0 series and will be launching Oct 18th. Cayman will be sometime in mid-late Nov.

[Iconyu]

Should be, at a minimum, very close in performance. New rumor is that Barts is actually performing like an early rumor, around GTX480 levels.

From the info I have heard, Barts IS the 68x0 series and will be launching Oct 18th. Cayman will be sometime in mid-late Nov.

That would totally ruin nvidias day. Seriously, with that few shaders taking on the 480gtx is a bit of s stretch it would make Cayman near invincible.

[kuroikenshi]

oh... antilles????

[Nintendork]

Just per se Antilles sound like the big daddy.

[spursindonesia]

i don't freakin believe Bart will be in GTX 480 performance territory, even the fanATic in me isn't crazy enough to put much weight on that rumor, my nVidiot side still instills some dose of reality in my mind. Bart i@ ~HD 5850 performance is very much believable, Bart getting close and at times surpass HD 5870 is still somewhat probable, but Bart the 260-280 mm^2 chip surpassing GTX 480 the 530 mm^2 behemoth ?? That would be totally earth shattering in 3D graphic.

[halfwaythere]

Barts Xt will probably get close to the 5870 in overclocking since the stock clocks are a tad small. I think AMD will leave a decent oc headroom since the GF104 clocks really nice even at stock volts and cooling.

[Calmatory]

As I posted in the other thread... Note, that all the numbers assume the actual 4+1 -> 4D transition with only 98.5 % loss in performance.

Barts Pro will have some 960 SPs. At 850 MHz core clock it can achieve maximum theoretical of ~2 TFLOPS, or something along those lines.
Barts XT will have some 1280 SPs. At 850 MHz core clock it can achieve maximum theoretical of 2.67 TFLOPS, or something along those lines.
Barts XT would need 863 MHz core clock to surpass Cypress XT's 2.72 TFLOPS shader performance.
Barts XT would need 952 MHz core clock to achieve 3 TFLOPS.
Cayman with 1920 SPs at 850 will hit 4 TFLOPS. Give it 952 MHz for core and it will go 4.5, and 1058 and it will go 5 TFLOPS. Single chip. At 40 nm.

But again, note that real world performance does not scale linearly with shader power, so the gains won't be as big as the raw numbers would suggest. Or, at least the performance boost is not because of raw shader performance. We do not know if they've done changes to tesselation unit. Since people say that it's the weak spot in DX11, I'd guess they would focus some effort there too, bringing the performance up even more.

Barts XT with 3 TFLOPS(should be able to get it to ~950 MHz, no?) + more advanced tesselation unit = GTX 480 will cry.

Oh, and chip sizes should go up a bit, but Barts XT should be smaller than Cypress XT etc. AMD's already unbeatable Perf/mm² should be up by some 20-25 %.

[flopper]

As I posted in the other thread... Note, that all the numbers assume the actual 4+1 -> 4D transition with only 98.5 % loss in performance.

Barts Pro will have some 960 SPs. At 850 MHz core clock it can achieve maximum theoretical of ~2 TFLOPS, or something along those lines.
Barts XT will have some 1280 SPs. At 850 MHz core clock it can achieve maximum theoretical of 2.67 TFLOPS, or something along those lines.
Barts XT would need 863 MHz core clock to surpass Cypress XT's 2.72 TFLOPS shader performance.
Barts XT would need 952 MHz core clock to achieve 3 TFLOPS.
Cayman with 1920 SPs at 850 will hit 4 TFLOPS. Give it 952 MHz for core and it will go 4.5, and 1058 and it will go 5 TFLOPS. Single chip. At 40 nm.

But again, note that real world performance does not scale linearly with shader power, so the gains won't be as big as the raw numbers would suggest. Or, at least the performance boost is not because of raw shader performance. We do not know if they've done changes to tesselation unit. Since people say that it's the weak spot in DX11, I'd guess they would focus some effort there too, bringing the performance up even more.

Barts XT with 3 TFLOPS(should be able to get it to ~950 MHz, no?) + more advanced tesselation unit = GTX 480 will cry.

Oh, and chip sizes should go up a bit, but Barts XT should be smaller than Cypress XT etc. AMD's already unbeatable Perf/mm² should be up by some 20-25 %.

Tesselation is not a weak spot in amd current 5000 serie, it has enough and more power for todays games and next years games.
the only time it is low on power is extreme tesselation in a syntetic software benchmark, and ask yourself, do we see that level of tessealtion in any games ever?
Maybe 10 years from now..maybe.
so the 5850 I got still going good for 10 years without an issue.

But I do hope they got some good processing power in the upcoming series,
my eyefinity setup requries more.

If I played with one screen, then not.

[spicypixel]

10 years? that's mighty optimistic...

[spursindonesia]

yes Evergreen tesselation capability isn't horrific, but it must be admitted that the tesselation performance doesn't scale well as the architecture expanded, such as from Juniper to Cypress, because of one size fits all philosophy. Now, nVidia's tesselation approach is rather excessive for todays games, but one must acknowledge how well it scales with its distributed tesselation processing & extraordinary strong + multiplied geometry performance. With the latest rumor of RPE approach, i think ATi is somehow following nVidia's method (though perhaps only in philosophy, the detail must be quite different considering the overall architecture differences), let's say strengthening tesselation & especially geometry performances without going overboard & excessive.

[demonkevy666]

Tesselation is not a weak spot in amd current 5000 serie, it has enough and more power for todays games and next years games.
the only time it is low on power is extreme tesselation in a syntetic software benchmark, and ask yourself, do we see that level of tessealtion in any games ever?
Maybe 10 years from now..maybe.
so the 5850 I got still going good for 10 years without an issue.

But I do hope they got some good processing power in the upcoming series,
my eyefinity setup requries more.

If I played with one screen, then not.

yes we do Merto 2033. 1 game at least. it's like the crysis of DX11 lol

the biggest problem for ATI is that the tessellation units are = in terms of performance from 5670 = 5770 = 5870.
crossfire over comes this problem by using more units together. when your limited by number of cards that doesn't help
From what I've seen the 6,000 series are showing is the unit is better, or more of them in the single gpu. vs 5000.

[Newblar]

if the 5770 is this big, imagine a 6870 on 40nm... unless 5770 is aiming to be the new 8800gt and compete with 460gtx
if 5870 to 6870 gets the same sp increase, it would go against everything amd has been going for the past couple years

[Sn0wm@n]

its all rumors right now ... and amd might be sending different news to see who's leaking stuff .. so until i have legit numbers i wont believe anything ...

[Olivon]

Maybe 10 years from now..maybe.
so the 5850 I got still going good for 10 years without an issue.

Have you ever seen a graphics card which allow to play games for 10 years ? Seriously ?

[-Boris-]

Have you ever seen a graphics card which allow to play games for 10 years ? Seriously ?

Wasn't his point that the level of tesselation in synthetic benches wouldn't be seen in real world games in 10 years? It was from a tesselation perspective he thought it would do 10 years.