AMD Cayman info (or rumor)

[cold2010]

http://bbs.expreview.com/viewthread....om=recommend_f

[LightSpeed]

http://bbs.expreview.com/viewthread....om=recommend_f

Fake. 20 SIMD engines dont make 1600 streaming processors. Zz

edit: Damn, I thought they were talking about the 6850, now I see that its the 5870. Its fine then

[saaya]

http://bbs.expreview.com/viewthread....om=recommend_f

wow... thats dissapointing...
6000 is 10% more efficient/sp/clock and only has 320sps more... which is a mere 17% more than 5000...

so at the same clocks (and i doubt itll come clocked notably higher) 6970 will only be 27% faster than a 5870... and thats an UP TO... if geometry is limiting it will be more, but... when is that actually the case besides tessellation benchmarks?

6970 is definitely slower than the 580 then...
not much, but 5-10% on average i guess...

[Dimitriman]

wow... thats dissapointing...
6000 is 10% more efficient/sp/clock and only has 320sps more... which is a mere 17% more than 5000...

so at the same clocks (and i doubt itll come clocked notably higher) 6970 will only be 27% faster than a 5870... and thats an UP TO... if geometry is limiting it will be more, but... when is that actually the case besides tessellation benchmarks?

6970 is definitely slower than the 580 then...
not much, but 5-10% on average i guess...

Not really Saaya its 4d shaders therefore its 480x4 for Cayman vs 320x5 for Cypress = 50% increase.

I would estimate performance will be around 40% higher on average with even more notable differences in tesselation intensive benchmarks.

[saaya]

Not really Saaya its 4d shaders therefore its 480x4 for Cayman vs 320x5 for Cypress = 50% increase.

I would estimate performance will be around 40% higher on average with even more notable differences in tesselation intensive benchmarks.

why estimate if the 6000 series is already out and has been benched by dozens of sites? compare reviews and you will see that sp for sp the 6000 series is ~10% faster than the 5000 series.

Well can't say if this slide is fake or not, but I guess you completely overlooked 50% more SIMD engines than HD5870 and double poly/clock.

On paper at least it should eat GTX580 alive.

i dont know how much the increase in poly/clock will help... but i doubt itll help much besides in tesselation benchmarks and new games that come out in 2011... and there itll only matter with high tesselation settings and i have yet to see a game or let alone demo where tesselation actually makes me want to upgrade :P
sure, it MIGHT be faster than the 580, but from what ive seen so far, theres nothing hinting at it... if the 1920sp number is true, it will trade blows with the 580 at best, and probably be a tad slower.

You didn't read about Barts or you're ignoring it.
Barts 6870: 1120SP with full efficiency
Cypress 5870: 1600SP but only 1280SP being used most of the time
Cayman is not a 5D thing. Let's make a guess

Cypress 5870: 1280SP effectively used
Cayman 6870: 1920SP fully used (or 2400SP if we apply Cypress SP inefficiency)

lets make a guess... OR... OOOOORRR we look at real world benchmarks
go and continue doing your 4d 5d math all you want, actual performance boost of 4d over 5d is around 10%, sp for sp at the same clock

You disappoint me.

which means you have high expectations and think a lot of me? thanks
i shant be dissapointing you

Try not just counting on SPs amount, but also the SIMDs. Not saying that AMD mArch has been limited in math calculating or shader power, but with this quite revolutionized, new mArch, it might offer better utilization & upped efficiency compared to the current one applied in Evergreen family & Barts chips.

revolutionized? so far it sounds more like a tweaked rv800?

Say 30 SIMDs are divided into 2 "GPC", each with its own setup & rasteriser engine, then each "GPC" with 15 SIMD will be divided into 3 arrays of 5 SIMDs. Each array has 16 TMUs totalling 48 TMUs/"GPC". As a whole, 2 "GPC" with 48 TMUs each will make a 96 TMUs chip. Just a very, very raw speculation of mine, please don't take it too seriously.

why would they only do that for the 6900 series though and not 6800? sure, 6900 might be a different arch, but... that would be a bit weird wouldnt it? would be cool, but its more wishful thinking and looking for a reason HOW the 6900 could beat the 580 if you ask me :P

id love it... i just dont think its very realistic

Also, Nvidia is apparently readying a dual GF100 card, so AMD will be facing competition.

dual gpu card formula:
if ( perf/W perfMAX - perf/W perfMIN ) > ( perf/W SLI - perf/W single) release

basically, if you can get more perf out of 300W by using 2 gpus, then yeah... but sli doesnt scale that well, so using two gpus is always slower than using one gpu at higher clocks.

[japamd]

why estimate if the 6000 series is already out and has been benched by dozens of sites? compare reviews and you will see that sp for sp the 6000 series is ~10% faster than the 5000 series.

But Barts parts are still VLIW 5, no one knows yet how good (or bad) the VLIW 4 will perform. That 10% improvement seen on benchmarks are due to internal optimizations (Techreport states AMD resized some queue and buffers).

Having twice the triangle setup rate/clock and 50% more ALU power, plus the internal optimizations seen on Barts, those improvements alone should give a nice boost over Cypress IMO.

[Mechanical Man]

lets make a guess... OR... OOOOORRR we look at real world benchmarks
go and continue doing your 4d 5d math all you want, actual performance boost of 4d over 5d is around 10%, sp for sp at the same clock

Please show those real world tests showing 4sp config performance. All that can be said based on Barts is that performance is atleast 10% sp per sp as othwise there would be no point doing that big arch change(basically making new shaders).

why would they only do that for the 6900 series though and not 6800? sure, 6900 might be a different arch, but... that would be a bit weird wouldnt it? would be cool, but its more wishful thinking and looking for a reason HOW the 6900 could beat the 580 if you ask me :P
.

Maybe it has something to do with where the best efficiency lies at what core config. It could be that 5VLIW sweetspot is ~1,2k sp and after that it declines. And 4VLIW sweet spot is after that. Who knows? Im just saying that there is too much variables that are left unanswered that one could really say about beating or not beating 580GTX.

[SimBy]

wow... thats dissapointing...
6000 is 10% more efficient/sp/clock and only has 320sps more... which is a mere 17% more than 5000...

so at the same clocks (and i doubt itll come clocked notably higher) 6970 will only be 27% faster than a 5870... and thats an UP TO... if geometry is limiting it will be more, but... when is that actually the case besides tessellation benchmarks?

6970 is definitely slower than the 580 then...
not much, but 5-10% on average i guess...

Well can't say if this slide is fake or not, but I guess you completely overlooked 50% more SIMD engines than HD5870 and double poly/clock.

On paper at least it should eat GTX580 alive.

[trinibwoy]

Well can't say if this slide is fake or not, but I guess you completely overlooked 50% more SIMD engines than HD5870 and double poly/clock.

On paper at least it should eat GTX580 alive.

On paper the 5870 also eats the 580 alive. But yeah it's shaping up to be sufficiently faster than Cypress to be in >= 580 territory.

[Nintendork]

wow... thats dissapointing...
6000 is 10% more efficient/sp/clock and only has 320sps more... which is a mere 17% more than 5000...

so at the same clocks (and i doubt itll come clocked notably higher) 6970 will only be 27% faster than a 5870... and thats an UP TO... if geometry is limiting it will be more, but... when is that actually the case besides tessellation benchmarks?

6970 is definitely slower than the 580 then...
not much, but 5-10% on average i guess...

You didn't read about Barts or you're ignoring it.

Barts 6870: 1120SP with full efficiency
Cypress 5870: 1600SP but only 1280SP being used most of the time

Cayman is not a 5D thing. Let's make a guess

Cypress 5870: 1280SP effectively used
Cayman 6870: 1920SP fully used (or 2400SP if we apply Cypress SP inefficiency)


You disappoint me.

[spursindonesia]

Now your seeing what I was trying to get at before. Atleast if your not being sarcastic. If the gtx 580 was not released at all, pricing could be potentially worse too.

Not really, you were suggesting that AMD will put US$ 500 or more MSRP price tag on Cayman XT, i said US$ 479 at most, and that's on the record. What i was trying to say, the market condition (lack of supply and perhaps combined with healthy demand) will drive up the appropiate to aggressive official pricing from AMD into higher street price level, like what has been happening with Barts XT & Pro case.

Don't worry, if Cayman XT was indeed priced (MSRP) OFFICIALLY by AMD @US$ 499 or more, i'll keep my bet & stop posting for the next 6 months.

wow... thats dissapointing...
6000 is 10% more efficient/sp/clock and only has 320sps more... which is a mere 17% more than 5000...

so at the same clocks (and i doubt itll come clocked notably higher) 6970 will only be 27% faster than a 5870... and thats an UP TO... if geometry is limiting it will be more, but... when is that actually the case besides tessellation benchmarks?

6970 is definitely slower than the 580 then...
not much, but 5-10% on average i guess...

Try not just counting on SPs amount, but also the SIMDs. Not saying that AMD mArch has been limited in math calculating or shader power, but with this quite revolutionized, new mArch, it might offer better utilization & upped efficiency compared to the current one applied in Evergreen family & Barts chips.

96 TMUs

Decoupled TMUs groups from the SIMD's engine ?

I've heard rumor about AMD's version of nVidia's GPC (Graphic Processing Cluster).

Say 30 SIMDs are divided into 2 "GPC", each with its own setup & rasteriser engine, then each "GPC" with 15 SIMD will be divided into 3 arrays of 5 SIMDs. Each array has 16 TMUs totalling 48 TMUs/"GPC". As a whole, 2 "GPC" with 48 TMUs each will make a 96 TMUs chip. Just a very, very raw speculation of mine, please don't take it too seriously.

[-=DVS=-]

No one finds it interesting how each company declares somethings ?

Shooting big claiming "Fastest"



Conservative , claiming only "most advanced".

[japamd]

No one finds it interesting how each company declares somethings ?

Shooting big claiming "Fastest"



Conservative , claiming only "most advanced".

Semantics... AMD can't call Cayman their fastest GPU as they have Antilles on board, NVIDIA have nothing else so it's indeed their fastest GPU

[Heinz68]

I ask the same, since if we are going for the only metric we can extrapolate quickly, the Flops, its nowehere near 50%.. From 2.7 Teraflops to 3.0 Teraflops is only 11%....

How do you know the HD 6970 is 3 TeraFLOPS. It doesn't show on this slide which somebody claimed is a fake anyway.

http://bbs.expreview.com/viewthread....om=recommend_f

[zerazax]

How do you know the HD 6970 is 3 TeraFLOPS. It doesn't show on this slide which somebody claimed is a fake anyway.

6990 is 6gflops so Cayman xt is prolly at least half that

[Heinz68]

6990 is 6gflops so Cayman xt is prolly at least half that

At least half is not neceseraly 3TeraFLOPS which Pica084 was using in his calculation.

The HD 5970 has "Processing power (single precision): 4.64 TeraFLOPS" and the HD 5870 is 2.7TeraFLOPS. LINK

In order to get within the 300W load power AMD needed to reduce the clock and voltage on the 5970. Since the HD 6990 slide shows "Load/Idle board power 300W/30W" some similar adjustment was most likely done.

HD 5970 performs about same as HD 5850 Crossfire unless overclocked to at least the HD 5870 levels.

[LightSpeed]

6990 is 6gflops so Cayman xt is prolly at least half that

At least, but im sure more. The 6970 will surely be clocked higher