AMD: R9xx Speculation

I see 5 output: 2 displayport, 1 hdmi, 2 dvi

This Eyefinity 5

 

No, HDMI and one DVI are shared.
They really done everything to reduce the pad-limit for Barts.

 

As with the previous generation, the HDMI is usually shared with one of the DVI.

edit: too slow

 

I still find it unlikely for Barts to have 1120 ALUs of 4D shaders , wavefront size would be horrible .

1120 = 17.5 X 64 ALU SIMD (Can't happen)
1120 = 14 X 80 ALU SIMD (Wavefront size is disastrous)

In fact if it had 1120 ALUs , then it wouldn't probably be much different than an overclocked HD 5830 with 32 functional ROPs and boosted memory frequency , that would be enough for the HD 5830 to even overtake the HD 5850.

So either ALUs are 1280 or 960 , those are the only ones that make sense right now .

 

Given the expected performance then 1280 it is.

 

awesome

Still looking rather forward to Cayman than Bart.

 

Hmm, and Fudo says 1120.

http://fudzilla.com/graphics/item/20525-radeon-hd-6870-has-1120-shaders

 

Or it could be the good old fashioned 1120 5D ALUs , and Cayman is the only new architecture this year .

These sites are just parroting each other , one site says 1280 , other sites say the same , then another site says 960 or 1120 , and the rest follows mindlessly .

 

DisplayPort can daisy chain monitors, so it really doesn't tell us anything.

Could very well be 6 monitors.

 

According to the leaked (supposedly real) slides, the DP's can indeed daisy chain up to 3 monitors each, which would mean max display count is either 6 (only DP's) or 8 (DP's daisy chains + hdmi/dvi + dvi)

 

Do you really expect 1120 to come close to 1600?

 

Actually, 1120@900 should be quite similar to 1440@725.

 

Overclock a HD5850 to the 5870 clockspeed and you won't see too much of a difference. The shader count of the HD5870 with 20 SIMDs engines appears to be a bit over the top. Obviously one can balance the architecture better. Clock 1120 SPs to 900MHz, maybe widen some bottlenecks and you will come close, I'm sure.

 

If scaling of processing units was optimized, HD 6870 (1120SPs@900MHz) could reach ~85% of D3D10/9 performance of HD 5870. At D3D11 they could fixed the tessellator "bug", that HD 6870 is >100% of HD 5870.
Combined you may get a similar performance at lower price and power consumption and some nice new features.

 

Why not ? 1440 at 17% lower frequency already came very close .

Not to mention you are basing you judgment on Vanatge scores only , real world performance could vary a lot , also Xtreme Vantage scores are closer to HD 5850 than HD 5870 .

My vote still goes for the 1280 ALU theory , heck even 960 sounds more sensible than 1120 .

EDIT : too late ..

 

"The first benchmark is Dirt 2, Ultra DX11 at 1920x1200, 8xAA, 16xAF"


"The benchmark is a a 3-screen [BattleForge] 5760x1080 benchmark, run at DX10.1 mode, comparing HD 6850 Crossfire to GTX 460 SLI. In numbers, the HD 6850 CF is 1.6x 460 1GB SLI and 3.57x 460 768MB SLI."


http://vr-zone.com/articles/amd-radeon-hd-6800-dirt-2-battleforge-benchmarks-leaked/10091.html

 

I rather go by the chart of big and little brother than all of the different SP combos at this point.

 

Nice, could be limited by amount of memory though (the GTX460 768MB almost certainly is), I'd like to see numbers for comparison with less AA or lower resolution.

Pointless due to memory limitation, though I guess it means at least it shows an advantage for radeons in memory limited situations (I think that was the case already for earlier generations though).

 

The difference betwen 960 at 7xx MHz and 1120 900 MHz is realy small . What would justify the price difference (and also aditional power conector, higher tdp)
It seems fillrate and bandwith is still king of the fps.

 

Even 800 could... or any other count in fact (though probably not clock/clock).

If you take "perf/flop" efficiency, Evergreen is clearly behind anything else, with only some rare scenarii where it edges R700 (Juniper/RV790).

nVidia manages to equal (and even come ahead of) Cypress with 50% less math throughput, 38% less texture fillrate and completely different pixel fillrate characteristics. GF104/106/108 show the same behaviour with some more texture fillrate (+33%) and less pixel fillrate (-29%).

What prevents AMD to achieve such a result? What size would Juniper become with enough "uncore" improvements to somewhat catch up with nVidia on that front? What if we add some more SIMDs (2 or 4) on top of that?


I bet some (potato) chips Barts is actually faster than Cypress and most leaks seen thus far are wrong/biased to show it worse than it is.