AMD: R7xx Speculation

[hoom]

It really looks convenient for a redundancy part, but it's somewhat too disproportional to me.

An alternative would be that the Red is the L1 & the Blue below is the LDS.
I was going to do that but Jawed seemed insistent on a 17th ALU thingie.

Then there is no coarse redundancy though. Could be ATIs fine grained redundancy is enough that they don't need to have such coarse redundancy?

Regarding the ALUs being smaller in RV770, could ATI have worked out how to provide the DX10 INT functionality within the FP hardware, enabling them to remove dedicated INT hardware?
It'd be a similar sort of change as removing point sampling from the TMUs (?) & R300 was doing DX7/8 INT on FP hardware so its not like we haven't seen that before.

 

[digitalwanderer]

Thanks AlphaWolf, much appreciate the help.

 

[John Reynolds]

They fixed it with the latest 1.1 or whatever patch, and IHVs should have profiles in their drivers already. ATi scales to 2 GPUs but not beyond, with the 1.1 patch and latest drivers, about nVidia I have no idea. And even prior to the patch with some renaming you could get it to scale.

I wish that were the case, because these two 4850s I've got installed bring GRID down to 1-3 fps once CrossFire is enabled. That's with the 8.6 hotfix drivers and GRID 1.1.

 

[AlphaWolf]

I wish that were the case, because these two 4850s I've got installed bring GRID down to 1-3 fps once CrossFire is enabled. That's with the 8.6 hotfix drivers and GRID 1.1.

maybe try renaming the .exe?

 

[mczak]

Regarding the ALUs being smaller in RV770, could ATI have worked out how to provide the DX10 INT functionality within the FP hardware, enabling them to remove dedicated INT hardware?

Pretty sure there's not really any dedicated INT hardware in R600. The "simple" alus likely just have an adder 32bits wide instead of 24 (adders are much cheaper than multipliers), and the "fat" alu is known to be a fp40 unit so it can do 32bit muls. Actually we know the rv770 gained even more abilities (I think the extremetech article mentioned it), on rv670 only the fat unit could do integer shift, but now all units can do it (then again - shifters are not exactly transistor-heavy to implement...)

 

[John Reynolds]

maybe try renaming the .exe?

I've tried a few different name changes, such as fear.exe or ut3.exe but no luck so far.

 

[AlexV]

I've tried a few different name changes, such as fear.exe or ut3.exe but no luck so far.

Just tried the demo, which doesn't even have the patch, and it worked great with CF...so dunno what to say.

 

[DJ12]

Someone at Rage3d spotted this. Supposedly 4870 already being sold in Hong Kong. Looks legit... oh, some basic 3dmark scores too.

Interesting, seems to be little better than the scores for the 4850.

http://www.ocxtreme.org/forumenus/showthread.php?p=44949

I know it's a better CPU in the 4850 test, but would/should it really make that much difference in a synthetic test like this?

I am guessing it's not a real 4870 or it's real benchmark score.

 

[swaaye]

I'm going to have to disagree here. Of all PC architectures, RSX was by far the best at the time for the size. Nothing based on G80 (which was too late anyway) would be any faster unless you heavily modified it. G84 is the closest in size and transistor count to G71 (remember that the former is 80 nm), but far slower.

8600 GTS will smack around members of the 7900 family rather frequently and (visually) look much better doing it. It actually performs similarly to X1950 Pro.
http://www.techreport.com/articles.x/12285/5

 

[hoom]

Actually we know the rv770 gained even more abilities (I think the extremetech article mentioned it), on rv670 only the fat unit could do integer shift, but now all units can do it

I see.
I guess R600/RV670 were often int-shift limited then?

 

[AlphaWolf]

I know it's a better CPU in the 4850 test, but would/should it really make that much difference in a synthetic test like this?

3dmark06 is cpu limited. I've seen 4850's nearer 15k on a 4ghz quad.

 

[ChronoReverse]

Okay, I've confirmed to myself on my 4850 that 8xAA is actually applied and incredibly cheap.

Anyone have a theory on what ATI did to make it so incredible?



(8xAA is actually faster than 4xAA+Wide/Narrow Tent for me).

 

[BRiT]

Can you run the D3D FSAA Tester tool and provide us some screenshots? (I'm not sure if there's a newer version of this). There's also another D3D FSAA tool by Tommti Systems.

 

[ChronoReverse]

8xAA
http://img78.imageshack.us/img78/6650/8xaavr4.jpg
4xAA
http://img131.imageshack.us/img131/1218/4xaasz7.jpg
2xAA
http://img78.imageshack.us/img78/8638/2xaagv8.jpg
NoAA
http://img113.imageshack.us/img113/9651/noaalx5.jpg

 

[hoom]

Anyone have a theory on what ATI did to make it so incredible?

4 AA samples per clock & shader resolve
Instead of having to loop several times through feature/rate limited fixed function hardware for the resolve, the shaders should be able to combine high numbers of AA samples in one pass. (at least, if I'm correctly understanding how that works...)

 

[CarstenS]

Supposedly, normal box-filtered AA is resolved in the ROP/RBE with RV770 and only the CFAA-modes get the (higher latency) shader-resolve.

That said, I suspect that they can loop their RBE pretty cheap, as was the case with X1K and earlier series, where 6x MSAA did not drop in performance much as along as VRAM did not overflow. Maybe Nvidia have to take another route as soon as they can't do single-pass (up to 4x) anymore? Second pass at the setup maybe?

 

[Lukfi]

I don't think they'd re-implement that. IMO, it's the sheer overkill power of the shader core.

 

[Tim]

8xAA
http://img78.imageshack.us/img78/6650/8xaavr4.jpg
4xAA
http://img131.imageshack.us/img131/1218/4xaasz7.jpg
2xAA
http://img78.imageshack.us/img78/8638/2xaagv8.jpg
NoAA
http://img113.imageshack.us/img113/9651/noaalx5.jpg

I am more interested in the AF, I would have been very surprised if AA sample patterns had been changed - but the we know that the TMU has been changed quite a bit so the AF could be changed (for better or worse).

 

[Broken Hope]

Looks like someone got hold of a 4870

http://translate.google.com/translate?hl=en&u=http%3A%2F%2Fwww.hkepc.com%2Fforum%2Fviewthread.php%3Ftid%3D1007750%26extra%3Dpage%253D1

 

[Broken Hope]

Looks like someone got hold of a 4870

http://translate.google.com/translate?hl=en&u=http%3A%2F%2Fwww.hkepc.com%2Fforum%2Fviewthread.php%3Ftid%3D1007750%26extra%3Dpage%253D1

Also do these seem real?