AMD: R8xx Speculation

[Pressure]

Fudzilla says that the 5870X2 will have a TDP of 376W and that ATI is trying to lower that.

He's a ......, sorry, where are my manners, no need for name-calling.

Even the Radeon HD 4870X2 (TDP 286 Watt) does not consume twice that of the Radeon HD 4870 1GB (TDP 160 Watt).

They will of course be binning the RV870 chips along with optimizing circuitry to reduce the overall Thermal Design Power.

 

[ferro]

Actually the claim shifted between then and now. The orginal claim was RV770, now its moved to RV790...

They can't be bothered with little details like that. It's 60% faster than something, and therefore it's "just like Fudo predicted a few weeks back".

 

[Creig]

They can't be bothered with little details like that. It's 60% faster than something, and therefore it's "just like Fudo predicted a few weeks back".

Shhhhhh! You're giving away his secret!

 

[Jawed]

As long as the attributes per vertex is fixed, (for one draw call) does it matter what is the upper limit?

It matters when you're trying to make the hardware do this. For example D3D10 was limited to 16 when it was originally intended to be 32, it seems.

I don't know what your point was, to be honest. The hardware takes multiple clocks to do this, as I would expect it would take TS multiple clocks to produce all the new vertices for a single patch.

Both situations: attribute rasterisation and naked vertex generation, both take a variable number of clocks to run, based on input.

Jawed

 

[Jawed]

I compared 512MB models to avoid influence of different VRAM capacity (1680*1050, AA4x/AF16x, computerbase performancerating). HD4870 1GB is significantly faster than HD4770 512MB (up-to 40%) but that's due its VRAM capacity, not its bandwidth... HD4770 1GB results would be very interesting, I think it could outperform HD4850 512MB.

1680x1050 shouldn't be enough pixels in any game to exhaust 512MB of memory with 4xMSAA. (Well, OK, Arma 2 could, I imagine.)

http://www.computerbase.de/artikel/...770/19/#abschnitt_performancerating_qualitaet

Hmm, those results are definitely very different from my long-standing understanding of RV770 performance. Dunno whether to trust them or to attribute that to driver progress for RV770 since June 2008.

It's showing 17% advantage on RV770 with 1GB versus 512MB at 1680 4xAA/16xAF. I really haven't a clue what's going on there. More pertinently, only 14% advantage for HD4870 over HD4850 really looks bad considering the extra bandwidth.

Jawed

 

[rpg.314]

I don't know what your point was, to be honest. The hardware takes multiple clocks to do this, as I would expect it would take TS multiple clocks to produce all the new vertices for a single patch.

I read your earlier posts again and I now understand what you meant.

Both situations: attribute rasterisation and naked vertex generation, both take a variable number of clocks to run, based on input.

Which is why I think that there is ff hw (rasterizer like) for this, and not via a kernel.

 

[Jawed]

Which is why I think that there is ff hw (rasterizer like) for this, and not via a kernel.

But the fixed function interpolator unit has just been deleted it seems (there is no "Interpolators" block).

Now, sure, this unit may have been moved inside of TS for all I know. But Marco is saying that attribute rasterisation is a function of the main ALUs...

Jawed

 

[fellix]

Hallelujah!

 

[leoneazzurro]

So all the rumors about better AF was true, if that is a Cypress shot..

 

[neliz]

Hallelujah!

I that fo rizzle? holly shizzle!

 

[fellix]

The shot is a first hand test result, I can assure you.

Probably I could get some AA-pattern samples, but I'm not promising.

 

[Jawed]

But can you see a difference in any games?

Jawed

 

[Cypher]

But can you see a difference in any games?

Jawed

While I'm super-psyched about true angle-invariance in AF, that excitement is instantly quelled by the fact that there'll probably be no difference. Aliasing due to crap AF has been trounced by the enormous amount of aliasing we have due to materials doing non-linear calculations. It's still a nice step forward, though.

 

[nAo]

It was about time

 

[neliz]

The shot is a first hand test result, I can assure you.

Probably I could get some AA-pattern samples, but I'm not promising.

Please do, I want to hear Jawed drop from his comfortable horse-hair filled, silver lined, desk chair!

 

[elsence]

1680x1050 shouldn't be enough pixels in any game to exhaust 512MB of memory with 4xMSAA. (Well, OK, Arma 2 could, I imagine.)

http://www.computerbase.de/artikel/...770/19/#abschnitt_performancerating_qualitaet

Hmm, those results are definitely very different from my long-standing understanding of RV770 performance. Dunno whether to trust them or to attribute that to driver progress for RV770 since June 2008.

It's showing 17% advantage on RV770 with 1GB versus 512MB at 1680 4xAA/16xAF. I really haven't a clue what's going on there. More pertinently, only 14% advantage for HD4870 over HD4850 really looks bad considering the extra bandwidth.

Jawed

There are games (very few for now) that will show a hit even in 1280X1024.
For example Far Cry 2.
The 1GB ver is 10%-15% faster than the 512MB ver in 1280X1024
15%-20% faster than the 512MB ver in 1680X1050
20-25% faster than the 512MB ver in 1920X1200

The above are average fps.

The real problem are the minimum fps. (the difference is way higher than in the average fps)
(the low minimum fps are making the games feel much slower than what the average fps suggest...)

 

[trinibwoy]

But can you see a difference in any games?

Well ATI's take was that you couldn't see a difference between G80 and R600. They've probably changed their minds now though and it'll be like night and day between GT200 and RV870!

 

[Andrew Lauritzen]

Yay for AF! I'd wager that it'll take a fairly trained eye to see any difference in games from G80 or similar, but that doesn't detract from the excitement of finally getting to this stage

 

[fellix]

Comparison animation with G80's 16xAF HQ pattern:

 

[Sound_Card]

Well ATI's take was that you couldn't see a difference between G80 and R600. They've probably changed their minds now though and it'll be like night and day between GT200 and RV870!

That was not their take, that was the fan groups take. In seriously, I have never seen the difference. No hardware out there produces crap AF like GF7 anymore. Maybe RV870 is more angle independent then G80+?