AMD: R7xx Speculation

[Wirmish]

A bag of holding?

No, I already explain how AMD can pack 800 SP in this small chip -> here

 

[randomhack]

480 SPs .. not 800

http://www.tgdaily.com/html_tmp/content-view-37950-135.html

Amazon posted 4850 on the site accidentally. 480 SP @ 625 MHz.

 

[AlphaWolf]

http://www.tgdaily.com/html_tmp/content-view-37950-135.html

Amazon posted 4850 on the site accidentally. 480 SP @ 625 MHz.

and I can provide you with another link of a retailer showing 800. Amazon posting it doesn't make it true.

 

[Shtal]

http://www.tgdaily.com/html_tmp/content-view-37950-135.html

Amazon posted 4850 on the site accidentally. 480 SP @ 625 MHz.

That is for RV7000 480SP's, but not for RV770.

 

[Pete]

I'm pretty sure they just post whatever info they're given without vetting it, and that info is wrong in this case. The 3DMark perlin test scores confirm 800 SPs at 625MHz.

 

[no-X]

Maybe other numbers could be helpful, too...

3DMark 2006

fillrate - single texturing:
HD4850: ~ 7300
HD3870: ~ 6500
RV670X2: ~ 11400

fillrate - multi texturing:
HD4850: ~ 19800
HD3870: ~ 12300
RV670X2: ~ 25300

pixel shader:
HD4850: ~ 680
HD3870: ~ 360
RV670X2: ~ 710

vertex shader - simple:
HD4850: ~ 292
HD3870: ~ 315
RV670X2: ~ ?

vertex shader - complex:
HD4850: ~ 160
HD3870: ~ 180
RV670X2: ~ 370

shader particles:
HD4850: ~ 160
HD3870: ~ 90
RV670X2: ~ 170

perlin noise:
HD4850: ~ 335
HD3870: ~ 175
RV670X2: ~ 355

 

[Lukfi]

=>no-X: Didn't you say we were not supposed to publish these?

 

[CarstenS]

Maybe other numbers could be helpful, too...

3DMark 2006
[…]

shader particles:
HD4850: ~ 160
HD3870: ~ 90
RV670X2: ~ 170

perlin noise:
HD4850: ~ 335
HD3870: ~ 175
RV670X2: ~ 355

Interestingly, shader particles seemed a bit bandwidth bound on R6x-generation, as HD2900 XT scored about 128 there. But maybe that was just some buffer overflowing and spilling into VRAM, which might have been alleviated with HD4800-series.

 

[no-X]

=>no-X: Didn't you say we were not supposed to publish these?

The time has come

 

[Entropy]

When analyzing the test data from the HD4850, I think its worth remembering that the RV770 may have been designed targeting GDDR5, and that some of these early data may be unusually bandwidth constrained.
It will be very interesting to compare the two memory options both in particular feature tests, and in applications.

 

[spajdr]

If i may add, i know its R7xx thread, but just for comparison i add Geforce 280GTX score

Maybe other numbers could be helpful, too...

3DMark 2006

fillrate - single texturing:
HD4850: ~ 7300
HD3870: ~ 6500
RV670X2: ~ 11400
GTX 280: ~ 14300
fillrate - multi texturing:
HD4850: ~ 19800
HD3870: ~ 12300
RV670X2: ~ 25300
GTX 280: ~ 40177
pixel shader:
HD4850: ~ 680
HD3870: ~ 360
RV670X2: ~ 710
GTX 280: ~ 845
vertex shader - simple:
HD4850: ~ 292
HD3870: ~ 315
RV670X2: ~ ?
GTX 280: ~ 246
vertex shader - complex:
HD4850: ~ 160
HD3870: ~ 180
RV670X2: ~ 370
GTX 280: ~ 119
shader particles:
HD4850: ~ 160
HD3870: ~ 90
RV670X2: ~ 170
GTX 280: 177
perlin noise:
HD4850: ~ 335
HD3870: ~ 175
RV670X2: ~ 355
GTX 280: ~ 300

 

[air_ii]

Benchmarks posted by jimmyz seem to confirm recent performance rumours. Only a few tests so far though...

http://www.xtremesystems.org/forums/showpost.php?p=3059795&postcount=3

 

[fellix]

RV770 vs. RV670 vs. G92 clean compilation:

Measure it for yourself.

 

[nexus_alpha]

So some of the cost advantages have been overblown then, slightly.

This thread has been very rah rah ATI, but bringing it back around, Nvidia still has the faster single chip and that's going to count for a lot when the dust settles. And GT200b especially, if they start upclocking and downpricing the thing. If the 260 starts hitting 350 and below points, it's going to start becoming the volume gamers choice.

Heaven knows when they will reach volume for at less then 100 good dies per wafer.

 

[air_ii]

RV770 vs. RV670 vs. G92 clean compilation:

Measure it for yourself.

Looks like roughly 44% larger. However a top shot rather than an angled and rotated shot would be much better for comparison, since you can't exactly tell when the side of the chip begins.

 

[fellix]

~272 mm² by my accounts, relating to RV670 size, on that picture.

 

[CarstenS]

Heaven knows when they will reach volume for at less then 100 good dies per wafer.

Seconded. Maybe, the die'd need another shrink for that to happen?

 

[Wirmish]

~272 mm² by my accounts, relating to RV670 size, on that picture.

So I was right -> 276 mm²

edit: 176 -> 276

 

[nexus_alpha]

So I was right -> 176 mm²

Sure maybe if you edit the number a little :smile:

 

[Silent_Buddha]

So I was right -> 176 mm²

I'd love to see a performance chip come out at 176 mm² in the present...

Regards,
SB