AMD: R8xx Speculation

[DegustatoR]

I strongly disagree to that; while CSAA is definitely better than 4x MSAA it's in no way as good across the scene as 8xMSAA.

It's not as good but it's quite often damn close and it's certainly better than MSAA 4x. So if you're providing MSAA 4x and MSAA 8x comparision you should include CSAA 16x numbers also as this is the AA with which most of GF owners will probably play newer games because for them going from CSAA 16x to MSAA 8x means loosing up to half of performance while getting pretty small quality increase.

Apples to apples isn't the same as to apples to fruits looking like apples

Just because they use the same number of subsamples ATI's and NV's MSAA isn't exactly apples to apples either.

 

[Kaotik]

True, but then where does CSAA fit in? It seems wrong to just ignore it.

Where does ATI's edge detect or tent modes fit in? Nowhere really, and neither do CSAA modes.

 

[fellix]

Edit:

It's FAKE!

 

[DegustatoR]

Where does ATI's edge detect or tent modes fit in? Nowhere really, and neither do CSAA modes.

Why, let's include ED CFAA too!
Last time I checked however even 12x ED was seriously slower than 8x MSAA on 4870X2.
So I don't think that these numbers will change anything.
Well, maybe it'll turn out that 24x ED on Cypress is about 8x MSAA / 16xQ CSAA speed on G200, which would be an interesting result for sure.

 

[Sound_Card]

Looks like a dual core.
If that matters any....

Honestly though, it's amazing how much leakage happens with AMD.

 

[homerdog]

Maybe CSAA 16x can be compared to ED 12x (?)

Where does ATI's edge detect or tent modes fit in? Nowhere really, and neither do CSAA modes.

ED doesn't fit in because the perf penalty is outrageous. Tent modes are more a matter of personal preference.

ATI's cards are designed to be fast with 8xMSAA and NV's with 4xMSAA + CSAA. Sure I'd prefer 8xMSAA over 16xCSAA but the difference in IQ is not that great while the difference in performance certainly is on NV hardware.

Meh, sorry for the OT.

 

[Slappi]

Slappi: ATi numbers were around zero for the last quarters. nVidia and AMD (CPU div.) are in much worse situation. -$12M/quarter is nothing compared to the other two companies, esp. in these days.

Not true.

NVDA makes a profit. They have 1.5 Billion dollars and no debt.

AMD/ATI Loses about $250 million or so a quarter and are 4 billion dollars in debt.

NVDA is in a position of strong financials. One of the strongest in the chip business. AMD is on the verge of bankruptcy if they do not turn things around within the next 2 years.

 

[psolord]

Question to all:

The discontinuous line in the 5850 vs GTX 285 graph, shows an average of 30% performance gain over the GTX 285.

So are the following calculations correct?

4870 512ΜΒ=30% faster than 4850 512ΜΒ
4870 1GB = 10% faster than 4870 512ΜΒ
4890 = 10% faster than 4870 1GB
GTX 285 = 10% faster than 4890
5850 = 30% faster than GTX 285
So 5850 = 1,0 Χ 1,3 Χ 1,1 Χ 1,1 Χ1,1 Χ1,3= 2,25 Χ 4850 = Holy Crap man!

 

[CJ]

AMD is on the verge of bankruptcy if they do not turn things around within the next 2 years.

Isn't this what doomsayers have been shouting for the last couple of years?

 

[Silent_Buddha]

ED 12x also usually looks significantly better than CSAA 16x and MSAA 8x. So the perf hit is somewhat justified.

Regards,
SB

 

[fellix]

 

[3dilettante]

There's already a doom and gloom thread for AMD.

I'd much rather speculate on the image that showed up earlier in the thread.
If real, there are some odd things about it that would be worth expounding upon.
Its resolution is such that any number of photoshop sins could be covered up, though.

 

[psolord]

Breath deeply...

http://img32.imageshack.us/img32/5794/09091421276f23be6536170.jpg

I think that with some more effort, there is a room for this thing to house at least 384-bit memory interface, looking at the perimeter features.

Does this mean that the 5870X2 could be a "quad core" chip? Does this mean that we are going to see better scaling this time?

 

[mczak]

It's not as good but it's quite often damn close and it's certainly better than MSAA 4x.

Well CSAA quality really ranges from msaa 4x equivalent to comparable to 8x msaa to even better than that.
But 8x msaa is better because you get similar amount of antialiasing everywhere. It's still interesting though due to the small performance hit.

Just because they use the same number of subsamples ATI's and NV's MSAA isn't exactly apples to apples either.

Last time I checked they were damn near identical. Sure they use different sampling pattern, but this doesn't have any impact on the amount of computation needed, and "overall" it shouldn't have any impact on quality neither, since the sparse sampling patterns are quite similar. It's only a driver thing anyway, since the sampling patterns are freely programmable (though I dunno maybe some of the patterns are patented somewhere wouldn't surprise me...).

 

[Kaotik]

They're obviously not in scale, but I had to just glue them together quickly, I'll try to do in scale version later this evening if no-one else has done one before

 

[apoppin]

It's not as good but it's quite often damn close and it's certainly better than MSAA 4x. So if you're providing MSAA 4x and MSAA 8x comparision you should include CSAA 16x numbers also as this is the AA with which most of GF owners will probably play newer games because for them going from CSAA 16x to MSAA 8x means loosing up to half of performance while getting pretty small quality increase.


Just because they use the same number of subsamples ATI's and NV's MSAA isn't exactly apples to apples either.

http://alienbabeltech.com/main/?p=10638

This is very recent and it is part 3 of a series of IQ comparisons that goes goes back a few months where the author compares Nvidia's IQ to ATi's - in every way. It's the best one i have seen in a couple of years.

 

[fellix]

Duh!
Is it me, or I can see 32 SIMD clusters in there on the die shot?!

 

[stevem]

Looks quite regular, like cache blocks.

 

[fellix]

Looks quite regular, like cache blocks.

Yup, the regularity is too uniform across the array's elements.

 

[AnarchX]

Architecture details:
http://translate.google.ch/translat...&extra=page%3D1&sl=zh-CN&tl=en&hl=de&ie=UTF-8