The general consensus is ~400mm² for Cayman. At 400mm² Cayman would be about 20% larger than Cypress, so I'd expect a bit higher price (320-350 for PRO)... The gap between HD6870 and HD6950 can be fixed later by HD6930, or by some premium-OC models of HD6870...
AMD: R9xx Speculation
- Thread starter Lukfi
- Start date
Hi;
I have a few questions about the MLAA AMD uses with the new drivers:
- do they take into account the information from the Z-buffer when calculating the morphological antialiasing already? [same as here: http://www.iryokufx.com/mlaa/]
- can it be combined with MSAA already?
- would it be possible to use MLAA only on "Z-buffer edges" to mimic MSAA at a lower cost?
- and/or would it be possible to use "heavy" MLAA on "Z-buffer-edges" and normal/light MLAA on all other edges/discontinuous parts of the framebuffer?
I have a few questions about the MLAA AMD uses with the new drivers:
- do they take into account the information from the Z-buffer when calculating the morphological antialiasing already? [same as here: http://www.iryokufx.com/mlaa/]
- can it be combined with MSAA already?
- would it be possible to use MLAA only on "Z-buffer edges" to mimic MSAA at a lower cost?
- and/or would it be possible to use "heavy" MLAA on "Z-buffer-edges" and normal/light MLAA on all other edges/discontinuous parts of the framebuffer?
I think a coverage sampling pre-pass could be used as a feed-back information to aid the morphological filter for better and more precise pixel blending. NV already has this in their CSAA modes, and in DX11 the shader kernels can access both the colour and depth sampled values.
The additional cost in video memory consumption should be negligible.
The additional cost in video memory consumption should be negligible.
I have split one easy answer into four more easily digestible answers! read below...
I hope these answers were clear, helpful and insightful!
If you have any comments on these answers, please send them through a PM tot his account.
Disclaimer:
I have no problems with anyone using my prefabbed answers to answer MLAA qustions in any further posts from here on, even on other bulletin boards or in normal, daily, mano-a-mano conversations.
If you are unsure which answer to use, flip a coin! Heads is "Yes, it's a post-process filter." Tails is "No, it's a post-process filter."
No it's a post process filter.
Yes, it's a post-process filter.
No, it's a post-process filter.
No, it's a post-process filter.
I hope these answers were clear, helpful and insightful!
If you have any comments on these answers, please send them through a PM tot his account.
Disclaimer:
I have no problems with anyone using my prefabbed answers to answer MLAA qustions in any further posts from here on, even on other bulletin boards or in normal, daily, mano-a-mano conversations.
If you are unsure which answer to use, flip a coin! Heads is "Yes, it's a post-process filter." Tails is "No, it's a post-process filter."
Last edited by a moderator:
Neliz is probably right about the MLAA filter selectable in the driver.
But depending on the game engine such a filter could access additional information (as the Z buffer) to improve the result. But this requires either the implementation by the developer itself or one would need profiles for this purpose in the driver.
But depending on the game engine such a filter could access additional information (as the Z buffer) to improve the result. But this requires either the implementation by the developer itself or one would need profiles for this purpose in the driver.
Indeed,
the application of MLAA by developers would be welcome but for now I fear we must live with this "patch"
Given how pushed this chip already looks on the HD6870, I've got some serious doubts about OC models which are worth it and significantly (more than 5%) faster...
Given the 4D shaders I think all bets on SIMD counts are off... If that's the only simd change (and wavefront size stays at 64) I'd certainly expect more than 20 simds (which would only be 1280 SPs) - maybe at least 24 (3x8? 2x12?) simds or something like that (which would apart from the 1536 SPs give 96 TMUs, unless there are some other big architecture changes which could well be). Well maybe earlier rumors about 1920SPs are true so even 30 simds.
As for the geometry scaling, no idea...
With a voltage tweak, I think 1GHz is doable. It might not be worth it compared to Cayman Pro, though, especially considering power-efficiency.
I've never, ever been right about these things, but here's my prediction:
32 SIMDs, perhaps divided into 4 parts with some kind of distributed geometry.
16 4-way shaders per SIMD, for a total of 2048 shaders, slightly short of a straight doubling from Barts, but more than a doubling if you count (4,5)-way units.
Unknown Soldier
Veteran
FYI
COD4 - 4xAA
COD4 - MLAA
I've gone back and looked at all my screenshots before I installed the new drivers and after. Before, all my screenshots have the fps at 125fps, lowest 123fps. After the new drivers were installed the screenshots I have show, 80, 114, 119, 98, 111, 107 fps.
The pics above show a lot of edges with 4xAA while with MLAA I see minimal edges and the picture actually looks like it might've been draw i.e. the objects look slightly integrated. I like
Did you see the techpowerup numbers wrt voltage scaling? They indeed achieved 1 Ghz. With 1.3V instead of 1.2V (and barely any further even with 1.5V). I don't know if that's really good enough to make it into a viable product. It needs to run not only at cold days and preferably with still acceptable noise levels
. Maybe if they do some binning.Oh that would be a quite serious chip indeed.
Did you see the techpowerup numbers wrt voltage scaling? They indeed achieved 1 Ghz. With 1.3V instead of 1.2V (and barely any further even with 1.5V). I don't know if that's really good enough to make it into a viable product. It needs to run not only at cold days and preferably with still acceptable noise levels
You mean this? http://www.techpowerup.com/reviews/HIS/Radeon_HD_6870/32.html I hadn't seen it, but it's not that bad. It seems to me that with some binning, and maybe 1.35~1.40V, 1GHz should be doable. Again, power-efficiency wouldn't be good, and there wouldn't be much overclocking headroom left. To be honest, that's probably not a card I would buy. And I think AMD would do well to keep the price gap between Barts XT and Cayman Pro fairly small. At worst, introducing a Cayman LE (HD 6930? 6890?) would work too.
Also, if I'm not mistaken, hot chips run faster, though they leak more. So summer days wouldn't be a problem as long as the cooling solution is appropriate.
Hey, where's the fun in speculating about reasonable things?
PS: yay! 666 posts!
Thanks!
I'm trying to run them combined as much as possible, not sure if that actually gives the best result..
I did notice some really weird sparkly sensation when text is placed over alpha blended vegetation. but it's only when it's light text over dark parts of the vegetation.. it's realy weird.. but it's the only way you know you're not multisampling anymore.
Thanks!
Well 4x8 (16 x vliw4) simds doesn't sound too unreasonable. Especially if the chip was really meant for 32nm and retargeted for 40nm (in fact considering past generations a downright small peak alu increase). Hard to tell though without any good leaks about what else could have changed (where are the leaks?)...Hey, where's the fun in speculating about reasonable things?
I believe they already are binning for the OC models.
They seemed to have binned the retail 6870 very tight, probably for TDP reasons, which is contrary to how some of the samples were clocking or maybe they forgot to mention the 20-28% increase in voltage.
They seemed to have binned the retail 6870 very tight, probably for TDP reasons, which is contrary to how some of the samples were clocking or maybe they forgot to mention the 20-28% increase in voltage.
Last edited by a moderator:
Linky please?
http://www.neogaf.com/forum/showpost.php?p=24004842&postcount=507
That got it to work for me finally after a lot of attempts at other things. You need to make sure you have the hotfix installed first though.
If not, try this - works with 4-series cards too apparently.
http://forums.guru3d.com/showpost.php?p=3749879&postcount=347
Last edited by a moderator:
My God, who needs faster cards when you can run 5040x1050 with 5770?
MLAA is especially handy for Eyefinity, wonder when AMD will get it working with DX10/11 games.Similar threads
