[chavvdarrr]

Originally Posted by mao5

got official 6xxx pdf

960?

[Mianca]

Originally Posted by gongo

I think they...should...its a HD6xxx....a new generation part...something we should expect...if it was a HD5x9x part then maybe i will keep my expectation down. I want to see a $199 killer part, AMD 6600GT/8800GT so to speak. AMD has been weak in $199 segment for ages.

It's all about (a) market situation and (b) production costs.

Both factors are different now.


Lets start with (a) market situation here:

When it was launched, HD 5770 basically "competed" against GTX 260, i.e. the GT200b salvage parts. GTX 260 proved a little faster in most benchmarks, but lacked DX11. In the end, Juniper XT launched @ a price tag a little above GTX 260. And it sold like hot cakes.

Consider Barts XT now. If the recent rumors are true, it will compete against GTX 470. So why should they price it (significantly) cheaper than that? A quick search @ newegg reveals that the cheaper versions of GTX 470 ship for $299 at the moment. I wouldn't be surprised if Nvidia dropped those prices a bit in the wake of AMDs new products - but Barts XT still shouldn't ship for less than $250. Actually, given some new features and improved tesselation performance, they could price it @ $279 and you could still call it a good deal given the current market situation.


Now look at (b) production costs:

Juniper used a new, smaller production process when compared to RV770. Early yields arguably weren't that great - but Juniper still was a lot smaller (~ 2/3 of RV770), while delivering about the same performance and adding DX11 as a killer-feature. Power savings where big enough to use a simpler PCB and simpler cooling solutions.

Now consider Barts XT. It might come with some considerable perf/mm2 improvements over the "old" Evergreen architecture, but given current performance rumors, I still doubt it will be a lot smaller than Cypress - at least nowhere near the 1/3 die size savings AMD managed when going from RV770 to Juniper, i.e. from 55nm to 40nm production. Maybe they'll save ~ 20% die space this time around (which would put Barts ~ 270mm2 - still a great achievement in pushing perf/mm2 on the same process node).

In the end (and that's the most important point - a lot more important than the comparision with Cypress), Barts most certainly will be a lot bigger than Juniper. It will consume a lot more power, need a more complex PCB etc. Selling such a card @ the same price as AMDs "old" (considerably cheaper-to-make) midrange parts isn't economically feasible - and they won't do that unless Nvidias "competition" forces them to adjust their pricing (which, sadly, won't be the case for the next few month to come).


All that being said ... don't forget about Barts Pro when looking for a ~$199 "killer" part

[Arnold Beckenbauer]

http://we.pcinlife.com/thread-1520616-1-1.html

[UniversalTruth]

Uploaded with ImageShack.us


http://we.pcinlife.com/thread-1520616-1-1.html


Looks pretty nice.

[Jawed]

Sadly, while that slide indicates that ROPs have doubled over Juniper, Z-rate per clock hasn't. ROPs match HD5850 for what I guess is HD6750. This would appear to indicate that HD6750 is clocked at 725MHz, same as HD5850.

40.x texture rate indicates 14 TMUs at 725MHz = 40.6GTexels/s.

Compute Performance appears to be 1.6x TFLOPS. Divided by 725MHz and 14 indicates VLIW-5. i.e. 1120 ALU lanes for HD6750. Perhaps 1280 for HD6770?

[fellix]

Originally Posted by Jawed

Sadly, while that slide indicates that ROPs have doubled over Juniper, Z-rate per clock hasn't.

Let's hope they've done something to improve the compression, at least.

[Triskaine]

One SIMD still has 80 SP's, but now with 4D ALU's, which means that wavefront size has increased to 80.

Why would they do that? 64 was already big, now Branch-Granularity will suck even more compared to GF100.

But at least the 6770 will soundly beat the 5850 with those specs, that's for sure.

[Jawed]

20 VLIW-4s per SIMD? That's very strange.

[Jawed]

Originally Posted by Triskaine

One SIMD still has 80 SP's, but now with 4D ALU's, which means that wavefront size has increased to 80.

Why would they do that? 64 was already big, now Branch-Granularity will suck even more compared to GF100.

80 is a really really nasty number for hardware thread size. If this is true then a whole pile of code that enjoys the "squareness" of 64 or 256 sized workgroups is going to be a nightmare. A total fucking bitch, in fact. Not to mention the mismatch in banks for 32KB of LDS, which doesn't divide into 20s in any meaningful way.

I don't believe this.

[Triskaine]

Originally Posted by Jawed

I don't believe this.

I don't want to believe this either, but that's the only possible way to interprete that slide.

[John021]

Originally Posted by Jawed

80 is a really really nasty number for hardware thread size. If this is true then a whole pile of code that enjoys the "squareness" of 64 or 256 sized workgroups is going to be a nightmare. A total fucking bitch, in fact. Not to mention the mismatch in banks for 32KB of LDS, which doesn't divide into 20s in any meaningful way.

I don't believe this.

Common people do not understand this.. what does it means in terms of performance, power or something else?

Do drivers for these cards will be a nightmare to make?

[PatrickL]

Who said theses slides were not another fake attempt ?

[CarstenS]

This doesn't look like a typical AMD slide either. At least not the ones I know from the past. They have mostly been structured with a liitle bit of information per slide and - very importantly - the improvements in each area highlighted.

[Arnold Beckenbauer]

So this picture says, HD6770 can give you 2304 GFLOPS. How is this possible?
...


Crap, something went wrong.

[mao5]

Originally Posted by UniversalTruth

Uploaded with ImageShack.us


http://we.pcinlife.com/thread-1520616-1-1.html


Looks pretty nice.

FAKE

[UniversalTruth]

Ok. I want the real one then.

[mao5]

Originally Posted by UniversalTruth

Ok. I want the real one then.

already in some posts

[Triskaine]

Originally Posted by Arnold Beckenbauer

So this picture says, HD6770 can give you 2304 GFLOPS. How is this possible?
320x4 gives you 1280 SPs. So 1280 * 2 FLOP (MADD) * 0,9 GHz is 2073,6 GFLOPS.

This says otherwise:

http://www.wolframalpha.com/input/?i=1280*2*0.9

[Man from Atlantis]

Originally Posted by mao5

FAKE

give the real one then

[no-X]

There's no mention about transistor count and power consumption... it's also untypical, that 2/3 of the slide are consumed by specifications of previous gen. products. This slide also mentions 5 Gbps modules for all products, despite HD5xxx slides always mentioned exact numbers (e.g. 4.8 Gbps for HD5870).

[mboeller]

Originally Posted by mao5

already in some posts

Where?

[MarkoIt]

It's probably another fake slide.. it's not enough polished to be real.
but if what mao5 is true, that the official pdf are around it won't be much before we get real leaks. It would be about time, launch is in less than 3 weeks!

[EduardoS]

Originally Posted by Triskaine

One SIMD still has 80 SP's, but now with 4D ALU's, which means that wavefront size has increased to 80.

Another fake, but 80*4 = 64*5, I think there is enough rumors to dismiss the 4D ALUs rumor, it's 5D again, at least for this generation.

And I'm not sad about this... But drivers could improve, it's taking too long to see a decent compiler for this.

[caveman-jim]

Originally Posted by MarkoIt

launch is in less than 3 weeks!

it is?!

[neliz]

Originally Posted by MarkoIt

It's probably another fake slide.. it's not enough polished to be real.
but if what mao5 is true, that the official pdf are around it won't be much before we get real leaks. It would be about time, launch is in less than 3 weeks!

I surely hope that's wrong.