AMD: R7xx Speculation

[DegustatoR]

Look at the Vantage Fill Rate scores posted on last page.

Thanks but no thanks.
I try not to make much sense out of any 3DMark benchmark -- they're too inaccurate and too 'optimized' in drivers to be reliable in any way.

Imo they gone 8TF or 2x Quad-TMUs per Textur Unit, which should increase fill-rate at relative low costs.

Even if they did (which i find pretty unlikely all things considered) that's only 2x TF rate or 2x TS rate while SP number increases by 2.5x which means that such RV770 will still be more TMU bound than even RV670.
What's even more important is that if they'd increase texturing rate by 2x while increasing shading rate by 2.5x you'd expect that RV770 would be ~2x faster than RV670 in real world benchmarks (aka games) but from what i'm seeing that is not the case -- RV770+GDDR3 looks to be more or less on par with G92 which is quite low if it has 2x texturing power and 2.5x shading power of RV670 which is ~25% slower than G92.
So i'll be on the 16 or 20 TMUs bandwagon for now, sorry =)

 

[w0mbat]

@Morgoth the Dark Enemy: According to AMD Shader = ALUs & 64 5D ALUs = 800 Stream Processors.

 

[Neo.Luxembourg]

Has anyone some Die (not packaging) pictures from the R600 or RV670 ?

The best I could found on google was this slide, but it's to small : http://benchmarkreviews.com/images/reviews/video_cards/Diamond_HD3870/DieSize_PowerUsage.jpg

 

[trinibwoy]

So i'll be on the 16 or 20 TMUs bandwagon for now, sorry =)

It's pretty hard to ignore the fact that the vantage fillrate score doubled vs the 16 TMU 3870. But yeah, I also find it strange that these massive unit increases across the board aren't showing up in the overall vantage score. A 32 TMU, 1 teraflop 4850 shouldn't be anywhere near 9800GTX territory.

 

[Wirmish]

October 6, 2006 : "The next generation could potentially have 96 shader processors (;-)) and will exceed half a teraflop ??? of computing power." - Dave Orton

vs

alexio (XS) -> RV770 = 800SPs, Don't ask for the source.

 

[Tchock]

Thanks but no thanks.
I try not to make much sense out of any 3DMark benchmark -- they're too inaccurate and too 'optimized' in drivers to be reliable in any way.


Even if they did (which i find pretty unlikely all things considered) that's only 2x TF rate or 2x TS rate while SP number increases by 2.5x which means that such RV770 will still be more TMU bound than even RV670.
What's even more important is that if they'd increase texturing rate by 2x while increasing shading rate by 2.5x you'd expect that RV770 would be ~2x faster than RV670 in real world benchmarks (aka games) but from what i'm seeing that is not the case -- RV770+GDDR3 looks to be more or less on par with G92 which is quite low if it has 2x texturing power and 2.5x shading power of RV670 which is ~25% slower than G92.
So i'll be on the 16 or 20 TMUs bandwagon for now, sorry =)

You obviously forgot that the 4850 does NOT have 2x the bandwidth of RV670. It has even less b/w than a 3870.

 

[Love_In_Rio]

You obviously forgot that the 4850 does NOT have 2x the bandwidth of RV670. It has even less b/w than a 3870.

But in a test of crysis at 800x600 without AA and in high settings bandwith doesn´t mind... ( ok, if the posted test is legit ).

 

[v_rr]

Thanks but no thanks.
I try not to make much sense out of any 3DMark benchmark -- they're too inaccurate and too 'optimized' in drivers to be reliable in any way.


Even if they did (which i find pretty unlikely all things considered) that's only 2x TF rate or 2x TS rate while SP number increases by 2.5x which means that such RV770 will still be more TMU bound than even RV670.
What's even more important is that if they'd increase texturing rate by 2x while increasing shading rate by 2.5x you'd expect that RV770 would be ~2x faster than RV670 in real world benchmarks (aka games) but from what i'm seeing that is not the case -- RV770+GDDR3 looks to be more or less on par with G92 which is quite low if it has 2x texturing power and 2.5x shading power of RV670 which is ~25% slower than G92.
So i'll be on the 16 or 20 TMUs bandwagon for now, sorry =)

Not true because you are talking about textures.
You need texturing but at some sort of level. If you put there to much texturing it won´t get you anywere more performance.
Think about that

 

[Jaaanosik]

October 6, 2006 : "The next generation could potentially have 96 shader processors (;-)) and will exceed half a teraflop ??? of computing power." - Dave Orton

vs

alexio (XS) -> RV770 = 800SPs, Don't ask for the source.

He was talking about R600. Right?

 

[mboeller]

Ka-ching!

Jawed

But the reported 1006 Gigatexels would be nearly as high as the 1070 GTexel of the G9800 GX2 and 3 times the GTexel-Rate of the 3870 (!!)

 

[mczak]

But in a test of crysis at 800x600 without AA and in high settings bandwith doesn´t mind... ( ok, if the posted test is legit ).

Sure of that? If this uses for instance lots of float (or even just uncompressed 32bit) textures, memory bandwidth could still be an issue.
And don't forget that even with 32 TF units that is still only half of what the 9800GTX has.
So it certainly seems possible in some situations that this card won't be faster than a 9800GTX (texture fillrate bound) or even 3870 (memory bandwidth bound) while outclassing both of them in other situations (alu bound - see the perlin noise test).
Actually, if this card got only 5 shader clusters (and thus probably batch size increased to 128 from 64), it might not even be really faster than a 3870 even if ALU bound, if all you've got is only small batches. Though it could have 10 clusters with the same batch size as rv670...

 

[Jawed]

But the reported 1006 Gigatexels would be nearly as high as the 1070 GTexel of the G9800 GX2 and 3 times the GTexel-Rate of the 3870 (!!)

647 as reported by 3DMark Vantage.

Jawed

 

[mczak]

But the reported 1006 Gigatexels would be nearly as high as the 1070 GTexel of the G9800 GX2 and 3 times the GTexel-Rate of the 3870 (!!)

No, the Teratexel number is from gpu-z and not comparable (probably just bogus anyway). Look at the vantage extreme texture fill rate (650GT/s) which is a bit more than twice that of the 3850 (not sure exactly why it's a bit more than twice rather than a bit less considering the clock differences, but it's hard to tell since the number obviously doesn't really measure pure texel fillrate).

edit: btw this vantage texel fill rate test is odd not only because of the huge numbers. A 9600GT gets beaten by a HD3850? Sure the GF9x chips can't realize their theoretical fillrate, but "normal" multitexture fillrate tests (like the one in 3dmark06) still puts it way ahead in this category: http://techreport.com/articles.x/14168/4 - what the heck does this really measure???

 

[Love_In_Rio]

Sure of that? If this uses for instance lots of float (or even just uncompressed 32bit) textures, memory bandwidth could still be an issue.
And don't forget that even with 32 TF units that is still only half of what the 9800GTX has.
So it certainly seems possible in some situations that this card won't be faster than a 9800GTX (texture fillrate bound) or even 3870 (memory bandwidth bound) while outclassing both of them in other situations (alu bound - see the perlin noise test).
Actually, if this card got only 5 shader clusters (and thus probably batch size increased to 128 from 64), it might not even be really faster than a 3870 even if ALU bound, if all you've got is only small batches. Though it could have 10 clusters with the same batch size as rv670...

If this card in its best scenario could even lose with 3870 i don´t understand what have been ATI engineers thinking about. I supposse that an improvement must take into account all the variables and i still hope this is the case.

 

[ShaidarHaran]

If this card in its best scenario could even lose with 3870 i don´t understand what have been ATI engineers thinking about. I supposse that an improvement must take into account all the variables and i still hope this is the case.

You misunderstand. mczak was referring to an absolute worst-case-scenario in which a 4850 may not outperform a 3850/3870. Best case scenario could be more than twice as fast.

 

[DegustatoR]

It's pretty hard to ignore the fact that the vantage fillrate score doubled vs the 16 TMU 3870. But yeah, I also find it strange that these massive unit increases across the board aren't showing up in the overall vantage score. A 32 TMU, 1 teraflop 4850 shouldn't be anywhere near 9800GTX territory.

That's what i'm thinking. As for the Vantage score -- we'll see how legit it is.
Don't you find a little bit strange that while everyone's talking about 32 TMUs -- bam! -- here's a screenshot with 2x texturing rate on the supposed RV770 board? =)

You obviously forgot that the 4850 does NOT have 2x the bandwidth of RV670. It has even less b/w than a 3870.

From what was leaked i'm seeing that 4850 is _less_ B/W limited than 9800GTX -- it tends to be more faster in hires/AA modes then it is in lores/noAA. That doesn't bode well with your "but it's B/W limited theory" -)

Not true because you are talking about textures.
You need texturing but at some sort of level. If you put there to much texturing it won´t get you anywere more performance.
Think about that

Thanks, i thought about many things -)
But _any increase_ in ALU/TMU ratio in RV770 while it's generally agreed that R6x0 was way too agressive in this regard would still be a mistake from where i'm standing. Something less ALUs and more TMUs -- even the same R6x0 ratio but more of everything -- would be a smarter thing to do IMHO.
...Unless of course ATI's done with games and is thinking about stream computing only from now on. Then those 800 SPs will be great.

 

[mczak]

You misunderstand. mczak was referring to an absolute worst-case-scenario in which a 4850 may not outperform a 3850/3870. Best case scenario could be more than twice as fast.

Exactly. And actually, such a worst-case-scenario would be the vantage color fill test (which uses a 64bit render target). I doubt it gets more bandwidth limited than that. For some reason though, the 4850 is still a bit faster than the 3870 (3.45 vs 3 GPix/s). Maybe improved color compression scheme?

 

[Wirmish]

He was talking about R600. Right?

Maybe you're right... but where are the 96 SP (5D) ?

2006: ATi said R600 = 480 SP, but at launch R600 = 320 SP.
2008: AMD said R770 = 800 SP, but at launch RV770 = 480 SP ?

 

[w0mbat]

He said it could be.

 

[Lukfi]

ATi was probably planning R650 as a 65nm 96 ALU chip, but the project was stopped in favor of dual RV670.