AMD Cayman info (or rumor)

[AffenJack]

info ist just bull, come on guys. We have already seen the tesselation performance on slides, it's 2times 5870 in most cases, but maybe you can tell me how they get 4 times better performance out of two times better tesselation performance?


I think i've to go to some forums and also write my performance estimates and write i've sources at amd, looks like most here would believe it.

[SpuTnicK]

He saying that 6990 faster then 5970 for about 20%, and in 3-4 times where tesselation is used.

Its 6970 not 6990. And yes, I can speak Russian, and translation is accurate enough to understand it.

Concerning trustworthiness of that info, maybe something is exaggerated, but IMHO not all is BS.

[Lanek]

info ist just bull, come on guys. We have already seen the tesselation performance on slides, it's 2times 5870 in most cases, but maybe you can tell me how they get 4 times better performance out of two times better tesselation performance?


I think i've to go to some forums and also write my performance estimates and write i've sources at amd, looks like most here would believe it.

I don't say it's true or not, but the big problem of the Cypress for Tesselation was one of data-flow: ( I will quote Beyond3D article ):

they try to keep some data in shared memory (as far as we can see, they try to keep HS-DS pairs resident on the same SIMD, with hull shaders being significantly more expensive then domain shaders) but data to and from the tessellator needs to go through the GDS. There's also the need to serialise access to the tessellator, since it's an unique resource,.

Ok now with 2 GPE ( one for each Group of SIMD ), this cycle can be executed in parrallel, speeding the process to at least 2-3x ...

But this not all... this remove too a bad bottleneck who was completely slow down Cypress in some particular tesselation case .
I quote again Beyond3D article:

This hints at Cypress having to go through a slow path in this particular case (or rather, it hints at the existence of a small dedicated cache/buffer for system values,
that gets thrashed with larger triangle sizes as the bottleneck shifts towards the last stage of rasterisation).

As it seems they have solve those problems, the result can end pretty high in some situation ... where the gpu no more face this.

The slide you have put, don't show how fast will be the 6970 vs 5870 in a game with tesselation, just how much faster is the 6900 for calculate 8-10-12-32 pixel/triangle vs Cypress .
A case where the performance of the cypress was completely killed by the problems find by Beyond3D, will see a bigger performance jump, as the gpu will now work with his full potential in those case and
not only cause the tesselation is 2-2.5x faster).
You can compare it of what happen in the case of a CPU bottlenecking your GPU, you decrease the resolution and the fps don't move, you OC your CPU, and the fps will now make a jump..

But i m right with you, i don't really believe the numbers given on this russian forum...

(excuse my english, not my first language )

[Pantsu]

20% over 5970 might be in the realm of possibility, but since those 2.5x and 3-4x tessellation performance was mentioned I'm more inclined to believe this is just some random bs. Of course those numbers are possible in certain scenarios, like when 1GB isn't enough, but then it doesn't really matter, the perf difference is meaningful in playable limits only.

My bet would be about the same perf as 5970 for 6970. Kinda like 6870 is almost the same perf as 5870.

[Shadov]

info ist just bull, come on guys. We have already seen the tesselation performance on slides, it's 2times 5870 in most cases, but maybe you can tell me how they get 4 times better performance out of two times better tesselation performance?


I think i've to go to some forums and also write my performance estimates and write i've sources at amd, looks like most here would believe it.

Do you see tesselation not utilized in games and underperform?

Do we need more tesselation performance in current gen cards?

Yes it's nice to have, but its just another marketing checkpoint that dosent really matter. Maybe next year when new consoles come out tesselation will get more utilization, but I still doubt that we will see anything comperable to Heaven extreme mode in the upcoming years, which again is only a syntetic test.

Furthermore when we get to that tesselation levels used in games todays graphic cards wont be enough to run them anyway. To sum up tesselation is currently not a buying decission factor IMHO.

[zerazax]

info ist just bull, come on guys. We have already seen the tesselation performance on slides, it's 2times 5870 in most cases, but maybe you can tell me how they get 4 times better performance out of two times better tesselation performance?

Like Lanek wrote, 2x better in tesselation synthetics doesn't mean it can only be 2x faster in real world games

For instance, consider games where the GTX580 is far faster than Cypress - we know that the 580 is typically 40% faster, but in some games where tesselation is maximized, the GTX580 is in playable fps ranges (say ~30fps) whereas Cypress is brought down to 10fps. It's not at all impossible to be 3-4x that for the 6970 if tesselation was what was brigning Cypress to its knees, and now tesselation is done differently and so the rest of the GPU can still process the rest of the game.

Doesn't mean the 6970's numbers aren't fake, or that the numbers are when the game is unplayable as is (4fps -> 16fps is 4x faster, but irrelevant in the real world), but 2x faster in tesselation doesn't mean in game can't be more than 2x faster since a GPU is doing a lot more than just tesselation in a game, and being free from doing tesselation may speed up the rest of the game as well

[flopper]

2 weeks to go.
start your engines.
ready to rumble.
Place your bets.
gloves off.

driver on its way out to partners, cards to follow.

[hurrdurr]

Like Lanek wrote, 2x better in tesselation synthetics doesn't mean it can only be 2x faster in real world games

For instance, consider games where the GTX580 is far faster than Cypress - we know that the 580 is typically 40% faster, but in some games where tesselation is maximized, the GTX580 is in playable fps ranges (say ~30fps) whereas Cypress is brought down to 10fps. It's not at all impossible to be 3-4x that for the 6970 if tesselation was what was brigning Cypress to its knees, and now tesselation is done differently and so the rest of the GPU can still process the rest of the game.

Doesn't mean the 6970's numbers aren't fake, or that the numbers are when the game is unplayable as is (4fps -> 16fps is 4x faster, but irrelevant in the real world), but 2x faster in tesselation doesn't mean in game can't be more than 2x faster since a GPU is doing a lot more than just tesselation in a game, and being free from doing tesselation may speed up the rest of the game as well

I think that example is invalid since 580's pure tessellation capability IS, in fact, a couple times that of Cypress'. A 10x tessellation advantage might not translate to games unless there is a pure tessellation bottleneck.

[zerazax]

I think that example is invalid since 580's pure tessellation capability IS, in fact, a couple times that of Cypress'. A 10x tessellation advantage might not translate to games unless there is a pure tessellation bottleneck.

But that's the focus on the example - it's that Cypress being worse can hit a bottleneck a lot earlier and may cripple the rest of the card's performance, whereas being faster in capability may make the card many times faster than the tessellation performance alone

[tbone8ty]

omg this thread is ridic, when will it end :p

[ripken204]

it will never end!
it's a vicious cycle of rumors and rants.
all we need is a single news release from AMD to end this madness.

[SpuTnicK]

No, actually it will. 7-13 December

[zalbard]

If it doesn't get delayed again.

[saaya]

great! 6900 is faster in tesselation! now show me some games that use tesselation properly and i might buy one, what do you say? there arent any? but... but... then whats the point?

[Chickenfeed]

I'd love to see a game thats a proper poster child for DX11 but as far as I'm concerned such a game has yet to surface. I'm mostly worried about performance in titles today personally. Sure it will be nice when more DX11 content makes its way to the market (summer next year perhaps... ) to have solid hardware available but I remember with both my 8800GTX and later on 8800GTS 512s that I never felt there was a DX10 title worth its salt in those cards lifetime ( ie one that didnt incure a large performance penalty or being worth the hit visually ) It wasn't until I had my 4870x2 that felt running most DX10 games were playable.

I'd love to see this 3 - 4 times increase but odds are this is best case market speak. Even 2x would make it competitive so anything beyond that is just gravy.

[Shadov]

great! 6900 is faster in tesselation! now show me some games that use tesselation properly and i might buy one, what do you say? there arent any? but... but... then whats the point?

Exactly what I was saying!

Current games hardly use the hardware that's already on the market including tessellation.

[SpuTnicK]

that's because almost all of them are made for consoles with its 2 (or more) year old hardware.
sorry for off-topic.

[s1nykuL]

I'd love to see a game thats a proper poster child for DX11 but as far as I'm concerned such a game has yet to surface. I'm mostly worried about performance in titles today personally. Sure it will be nice when more DX11 content makes its way to the market (summer next year perhaps... )

With the possible exception of Crysis 2 expect DX11 games when PC GPU's are at DX13/14 level and consoles have DX11 capable GPU's in them.

Generally speaking and with very few possible exceptions, (none come to mind at the moment) PC games are not developed any more. Console games are developed and then ported to the PC, sometimes a DX10/11 code path is tagged on as an after thought.

[SpuTnicK]

Back to topic.
2x HD6850 have 1920 SPs as well as rumored for HD 6970.
However, tests show, that 2xHD6850 are on a par with HD 5970.
So, we can conclude that in case HD 6970 has 1920SPs, it will perform as HD 5970, or even better(don't forget higher clocks )

[zalbard]

2x HD6850 have much higher VRAM bandwidth, though.
And we don't know for sure whether everything gets doubled or not.
It's all speculation...

[Dami3n]

Back to topic.
2x HD6850 have 1920 SPs as well as rumored for HD 6970.
However, tests show, that 2xHD6850 are on a par with HD 5970.
So, we can conclude that in case HD 6970 has 1920SPs, it will perform as HD 5970, or even better(don't forget higher clocks )

Not only clocks. Don´t forget new 4d architecture, more TMUs, 2 tesselation units, more VRAM.

[tajoh111]

Back to topic.
2x HD6850 have 1920 SPs as well as rumored for HD 6970.
However, tests show, that 2xHD6850 are on a par with HD 5970.
So, we can conclude that in case HD 6970 has 1920SPs, it will perform as HD 5970, or even better(don't forget higher clocks )

If it perform's like xfire 6850's. I am pretty sure they would have released it already. Paper launch or not, it would have put a pretty strong damper on the gtx 580 launch in addition to raise the 6xxx namesake, to increase sales of the 68xx cards even further.

Not only clocks. Don´t forget new 4d architecture, more TMUs, 2 tesselation units, more VRAM.

They are not going to double everything unless they want the card to consume 300 watts + be 500mm in size.

I think the 3-4 tesselation performance is bull. Synthetics show the greatest difference in tessellation performance. People rationalizing that Cayman could be 2-3 faster in synthetics and even faster in games by the vagueness in the rumors shows how misplaced their rationality is.

Synthetics are coded to specialize in showing maximum prowess of a certain aspects. Drivers are specifically coded for them to maximize performance, hence why Nvidia kept on bragging about unigine heaven.

We won't see faster tessellation in games compared to synthetics. Be rational people.

[SpuTnicK]

According to latest slides (from Nov. 22nd). 4d is similar to 5d performance. Number of ROP number has not changed. TMU number for HD 6850 is 48. HD 6970 will have more than 80 (according to slide.) which is almost double. Double the vram (2Gb), but less likely not memory bandwidth

[zerazax]

2x HD6850 have much higher VRAM bandwidth, though.
And we don't know for sure whether everything gets doubled or not.
It's all speculation...

That's only going to matter if the bandwidth is necessary

[Manicdan]

didnt the scaling of the 6870 depend mostly on bandwidth? cause it was given some cheap stuff, and i know we cant speculate on how 4D reacts to bandwidth based on 5D numbers.