AMD: R8xx Speculation

[psolord]

TechPowerUp has some more sneak-peak shoots of Evergreen card demoed at QuakeCon2009

Head over to this post ...

It looks nice :smile:
Red PCB and Black radiator

PS. One shot is focused at air exhaust and it looks like radiator itself is small and aluminum!

I must be reading something wrong. The 3dmark06 test shows 38000 score for the 4870X2 and mine is almost half that, with a 4Ghz Q9550! Did they use two 4870X2s in that test?

Don't mind the 6.6Ghz Phenom. Their SM2 and SM3 tests were also double mine!

 

[Lukfi]

Wasnt the ring buss there in R520 ?

Sort of, IIRC it was a combination of ring-bus (for memory reads) and crossbar (for memory writes) (hmm, does what I just said even make sense? :neutral: )

 

[Jawed]

He said conceptual change, not about change in specs.

The two go hand in hand, effectively, i.e. when a conceptual change happens what you know about the unit specifications and counts goes out the window. RV670->RV770 is minor in comparison, it would seem, if Eric's to be believed.

Jawed

 

[rpg.314]

We are set for an X2card made of 2MCM's then?

may be even3,

 

[neliz]

We are set for an X2card made of 2MCM's then?

may be even3,

There is no "X2"

 

[psolord]

There is no "X2"

Do we know anything about the naming scheme yet?

I really hope the codenames, Juniper, Cypress, Guns'n'Roses, remain exactly that, codenames and they will not find their way on retail boxes.

Oh the embarassment to answer, "ATI Radeon Hemlock", when asked what graphics card do you have!

 

[neliz]

Do we know anything about the naming scheme yet?

I really hope the codenames, Juniper, Cypress, Guns'n'Roses, remain exactly that, codenames and they will not find their way on retail boxes.

Oh the embarassment to answer, "ATI Radeon Hemlock", when asked what graphics card do you have!

Since we HAD X and HD already. We Have a new DX version. it's probably something embarrassingly "Radeon 5870 Windows 7 Edition"

Or mix Win7 with DX11 and we have the Radeon 7/11 870

 

[rpg.314]

There is no "X2"

Interesting. May be X3 or X4 then? I shudder to imagine the power requirements of a card that is an X4 card of 181mm2 chips, and that too on TSMC's 40 nm process.

 

[psolord]

Since we HAD X and HD already. We Have a new DX version. it's probably something embarrassingly "Radeon 5870 Windows 7 Edition"

Or mix Win7 with DX11 and we have the Radeon 7/11 870

Personally, i would be ok if they kept the naming scheme to something we all love and understand. Ie Radeons, 5650,5670,5850,5870,5870X2 or 5970 if they fancy.

DX11-200, DX11-300, DX11-400 would be fine too! But putting anything Windows 7 on the name it would be pretty much disgusting.

They can put all the Win7 logos they want on the box. Not the name!

 

[Humus]

Not seen this posted

http://www.legitreviews.com/article/1040/1/

Isn't the performance hit of High Definition Ambient Occlusion kinda high? Doesn't look like it would be really usable in games.

I would assume that's because they take an extreme number of samples in this demo. Normally HDAO is probably the cheapest of the SSAO-like techniques around.

 

[mboeller]

See the spreadsheet:

http://forum.beyond3d.com/showpost.php?p=1316031&postcount=1252

I have updated the spreadsheet since I originally posted it, so see worksheet "JA02". Note the comment attached to RV740's L2 unit count.

As you can see I have ~9mm² I can't account for in RV740.

If you have any ideas or want to make more precise measurements of units on RV770 - instead of that generic "control, setup, interpolation, tessellation etc" - then feel free! It's worth noting that I included some of the "shiny" border between blocks when measuring the sizes of various units. These shiny areas seem to be just metal routing without any logic, i.e. they are the spaces between blocks of logic or blocks of I/O. Also there's some seemingly dead space in the corners of the die.

Obviously Juniper is a wild guess with scalings that aren't particularly justifiable...

Jawed

Thank you very much for your spreadsheet. Your reasoning is far better than mine.

It seems you are correct with your justification that Juniper has only 640 Shaders. With the informations from the spreadsheet I was able to "simulate" the RV730 too and came within 5% of the original Die area. The only tweak I had to make was that I had to halve the area for "control, setup, interpolation, tessellation etc" and to use 4 clusters.

Damn.. I had hoped that Juniper would be something like RV740 x 1.5. So the improvements from ~P8000 for the RV740 to ~P9500 for Juniper "only" come from improvements of the design.

 

[dizietsma]

I've been having problems recently with graphics in that I lost all interest for 12 months, I even thought half way through Crysis Warhead "Why am I bothering to play this" so stopped.

So now I having to catchup on the 2 threads for ATI and nvidia as I have decided to leave Windows XP and DX9 land like millions others to travel to thae land of milk and honey that is Windows 7 and DX11. So obviously I will be getting ATi R8 or nvidia GT3 thingies.

I used to be diehard nvidia but my last card was the fruity HD 4850 so I am completely neutral this time. The ATI seems more appealing but maybe this is because more details are out and it is closer to release?

I'm hoping for the standard 40-50% increase over the last generation so I can hit me good ol' 1920x1200 ideal; is the general concensus that is what will happen? I assume IQ is going to be still on par which ever way.

I told you I was out of it for too long.

 

[Davros]

fruity HD 4850

!!!

 

[dizietsma]

!!!

A peach without being top banana.

 

[Ailuros]

There is no "X2"

3 digits are funkier than 4 and stop confusing the crowd please LOL

 

[Jawed]

Thank you very much for your spreadsheet. Your reasoning is far better than mine.

Yes, but I have no track record with future products

I was arguing that RV770 was 480 ALUs (and generally unexciting) right up until the evidence was too strong

It seems you are correct with your justification that Juniper has only 640 Shaders. With the informations from the spreadsheet I was able to "simulate" the RV730 too and came within 5% of the original Die area. The only tweak I had to make was that I had to halve the area for "control, setup, interpolation, tessellation etc" and to use 4 clusters.

Aha, I haven't tried that. I've been meaning to but forgot. This spreadsheet is much better than my older ones, in which I've tried all sorts of things. The GDDR3 interface on RV730 should be smaller - the difference in pins is probably only ~8 per 32-bits (e.g. 53 pins for GDDR3 versus 61 for GDDR5?) and the power levels should be lower, making the interface smaller again...

Damn.. I had hoped that Juniper would be something like RV740 x 1.5. So the improvements from ~P8000 for the RV740 to ~P9500 for Juniper "only" come from improvements of the design.

Vantage scores are useless.

I'm not counting any chickens before the eggs have hatched. AMD has gone on record saying this is the most radical change since R600...

Jawed

 

[DegustatoR]

So the improvements from ~P8000 for the RV740 to ~P9500 for Juniper "only" come from improvements of the design.

...Or higher frequencies...

 

[Ailuros]

That should be correct, but then on the other side even if there are significant efficiency changes in an architecture would Vantage be a good candidate to show any of them?

 

[neliz]

...Or higher frequencies...

I believe they call it "Pixie dust infused Dual Shaders" over at AMD.

 

[Forrest]

RV870 (Cypress)

SIMD : 20
Shader Clock : 850
Memory Clock : 1200
Bandwidth : 153Gbps
1024Gb GDDR5

enjoy