Everybody, give me your predictions. Foresee.
My predictions for R700/RV770:
RV770 (RV670 for comparison)
-400 SPs (320)
-> increased the number of shader-unit blocks from 8x40 to 10x40
-40/40/40 TextureAdressUnits/TMUs/FP32Filters (32/16/20)
-> doubled the number of TMUs per Texture Processor from 4 to 8, increased the number of FP32 filtering units per TP from 20 to 32, and increased the number of tex processors from 4 to 5, to keep the Sp-to-TP ratio intact.
also
-16 ROPs (improved)
-256-bit external memory interface, tweaked a bit + ability to make two chips share one memory pool.
R700-> 2x RV770, but with the ability to share the memory pool, which is why the memory interface can be called 512-bit wide in this case.
I won't make any predictions about clockspeeds, because probably even AMD themselves don't know them yet and will decide upon them depending on how exactly yields and power consumption turn out to be.
GT200-240 SPs
-> arranged in 10 arrays of 24 SPs each
-80 or 120 Texture Units
-> NordicHardware claimed it will have 120, but I find 80 more likely, because for 120, they would either have to f*** up the SP-array to TextureProcessor ratio (which is 8/8 on G92, so 10/15 on GT200 sounds a bit strange) or increase the number of TMUs per TextureProcessor from 8 to 12, but that would make little sense IMO, since the need for additional SPs grows faster than the need for additional TMUs and even G80/92 have a comfortable lead over the competition already, 120 would be a bit overkill IMO.
-32 ROPs for GTX, 28 for GTS
-512-bit memory interface for GTX, 448 for GTS
-1GB GDDR3 for GTX, 896 MB for GTS
About RV770, I'm aware that most rumors say 480SP/32TMU, but it is common knowledge that one of the main bottlenecks of R6xx was the bad SP-TMU ratio, if you look at the TMUs only and ignore the TAUs/FP32 filters then 480-32 might sound better than 320-16, but if my predictions that they increased the TMUs per TextureProcessor and not just the number of TextureProcessors are true, then 480-32 would actually
worsen the SP-to-Texture
Processor ratio, that's why I find 400-40 more likely. Besides, there are some other minor things that support my 400-40 theory. First, as I pointed out above it would kepp the SP-array-to-TextureProcessor-ratio intact, but would heavily improve the SP-to-TMU&FP32Filter-ratio, basically removing one of R6XXs main bottlenecks.
Then there's those chinese sites, who claim the chip will have 800 SPs. IF the rumored die-size/transistor-count are correct, that's impossible, I mean if it was possible to squeeze 480(!)SPs and 16 TMUs into only ~170million additional transistors, I'm sure AMD would have done that with RV670 or a single-chip R680 already.
What I think is, that those chinese guys simply mixed RV770 with R700 here. 2x400=800, plain and simple. Last but not least, I remember i read somewhere on VR-Zone a claim that when a source was asked if RV770 would feature doubled TMUs, it is claimed to have replied "more than that".
Also, I read another rumor on VR-Zone according to which an AMD guy has claimed that RV770 might be "up to 100% faster in some situations". If you look at benchmark figures in recent graphics cards reviews, you'll eventually see that in some games the G80/G92-cards are up to 70% faster with AA/AF enabled, and even up to 110% faster when DX10/HDR/AF are combined. I think it's these situations in which Rv770 will shine, at least compared to RV670, and especially in the latter example, more Texture Processors with more FP32 Filter units and more TMUs would help more than additional SPs.
The last thing that speaks for my theory is, in fact, Nvidia. Nvidias G8x/G9x are undoubtedly more popular and widespread amongst gamers, also there's Nvidias TWIMTBP program, which probably makes Nvidia more popular and widespread amongst game developers, too. The result is that many games get more optimised towards NVs 2-to-1 Shader-to-TMU ratio because they want the game to run well on the most widespread technology, and not towards AMDs 4-to-1 approach on the R600/RV670. But with 400 or actually 80 (I won't bother explaining, you know how AMD is counting them) SPs and 40 TMUs, AMD would switch to Nvidias 2-1 ratio - which might result in much better performance even in Nvidia-optimised games, simply because there's suddenly no difference in the ShaderUnit-TextureUnit ratio anymore.
Yeah i know, some of my deductions a pretty far-stretched and based on rumors that could turn out to be complete BS, but still,
+25% SPs,
+25% TextureProcessors containing 100% more TMUs and 60% more FP32-Filters each (resulting in 150% more TMUs and 100% more FP32-Filters in total),
+higher clockspeeds,
+faster memory and
+tweaked ROPs
could very well result in the rumored 40-60% more performance under normal circumstances and up to 100% in some situations that need lots of Texture -Mapping and -Filtering, they don't contradict the rumored die-size, they don't even contradict the 800SP rumor if you factor in R700... I dare to say the only problem is that my predictions contradict the popular 480SP-32TMU rumor. If the "AMD documents" that the 480-32 theory was based upon are real, my prediction is pointless, if they were fake, there's at least a chance that I'm right (or at least close to the truth).
