AMD: R7xx Speculation

[Jawed]

If ever there's a game to test on HD4870X2, it would seem to be Racedriver GRID. From the FAQ:

viii) Is ATI’s Crossfire or NVIDIA’s SLI supported?

Unfortunately, due to the working of GRID’s engine, systems which link two or more graphics cards (i.e. Crossfire or SLI) are not implicitly supported. This means you may not see a performance boost from having the second card, and you may even have trouble running the game. If GRID is failing to start properly and your system is SLI or Crossfire enabled, try disabling the technology in either NVIDIA’s or ATI’s control panel in Windows and see if this helps.

http://community.codemasters.com/forum/showthread.php?t=279124

Jawed

 

[Mintmaster]

Its a great chip compared to RSX thats for sure (at least in terms of its design and functionality) but thats more because RSX was pretty poor for its timeframe.

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.

Xenos stands out because it was designed from the ground up for a console. It has a high setup rate (not useful on the PC in 2005), can get away with EDRAM (128-bit bus no longer hurts perf), there's no need to build a scalable architecture (no wasted trannies on making 4 smaller units instead of one big one), etc.

What I mean is, Xenos is clearly a great design on paper, and it also comes packed with great functionality but the same can be said of R600. We mark R600's "greatness" down because it didn't perform as well as we expected. I'm just not seeing why we should assume Xenos is a superior implementation of the architecture when R600 came second and had time to learn from and refine the Xenos design.

First of all, you have to realize that R600's shortcomings are all related to perf/mm2, not the relatively useless metric of perf/pipeline.

R600 really is quite different. There's a radically different memory system (which they sort of ditched now) and far different shader structure (Xenos had one scheduler and one texture unit block, R600 had 4), texture filtering unit (R600 spent mostly useless trannies here), DX10.1 and PC related extras, and maybe going from 4D+1D to 5x1D w/integer originally took a lot more transistors than it should have.

Assuming Xenos is as efficient an implementation of that basic architecture as R770 seems a bit baseless to me.

True, but for pixels shader loads there's no reason to think it's closer to R600 than R580, and the latter's decent in terms of efficiency by even RV770's standards (but not unified or DX10, of course).

 

[Jawed]

High quality version is worthwhile

http://www.youtube.com/watch?v=ROAJMfeRGD4&fmt=18

Notice he says rendered with no polygons, even though he said the environment is CG assets? So its pre rendered right? He did say the city contains GI and photon maps, Not real time type stuff at the mo is it?

Yeah he seemed to be saying they have a variety of maps, which are static.

Rendering, though, appears to be voxel-based ray-tracing with wavelets used to compress the voxels - running in real time. The whole thing is ray traced, there is no triangle-z-buffer based rendering. Seems pretty unbelievable...

Jawed

 

[trinibwoy]

So what exactly is "static" in that scene?

 

[MfA]

You can "rasterize" voxels too (splatting).

 

[Jawed]

So what exactly is "static" in that scene?

Are you talking about objects or computed data?

Clearly most objects are static: the buildings (except for the bits that get knocked off), the cars (except for the taxi, those at the crossroads driving left-right and the car that escapes being crushed) and the road (except for the odd crater and debris) and the sky (which appears to be changing during 1:33-1:39 in the video, it looks like reflections of the sky on the cars are changing).

As to whether the "global illumination, photon maps, diffuse lighting" are static, well I dunno. Seems likely.

The way some transitions of rendering type occur, it looks like rendered results are cached for re-use in later frames.

Jawed

 

[Mintmaster]

That's clever!

I like that, particularly the solution to redundancy.

Jawed

Yup, I like it too. That's why I thought that similar-looking sliver next to the 4x10 array was redundancy, but then mczak pointed out that it looks a bit different. hoom's explanation tells us why - only the bottom half needs to look the same for the theory to work.

I guess instead of 25.2% or 28%, it's really 26.6% devoted to the ALUs. Eerily similar to the 26.5% that Arun concluded for GT200. :smile:

Now the only thing preventing a true apples to apples comparison of ALU density is the scheduling logic.

 

[Jawed]

So, if we take the "17 ALUs" that hoom has identified, we get 2.67% of the entire die. So that's 6.8mm2 for 120GFLOPs, or 17.6 GFLOPs/mm2, including the register file.

If we include what hoom reckons is the sequencer, then we have 8.7mm2 (3.4% of the die), which gives 13.8 GFLOPs/mm2.

Note that the most recent "big" picture of the die indicates that it is non-square, the height is 98.4% of the width...

Jawed

 

[fellix]

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

The fact is, that the whole ASIC is really very dense design and that last X-Ray micro-graph just don't have enough resolution to dive in more detailed dissection.

 

[digitalwanderer]

If ever there's a game to test on HD4870X2, it would seem to be Racedriver GRID. From the FAQ:



http://community.codemasters.com/forum/showthread.php?t=279124

Jawed

How did Guru3D get it working? Any ideas on what Hilbert did? I'm perma-banned from there so I can't ask and my buddy who wants to know says they've disabled registration.

 

[Silent_Buddha]

Amazing I remember the first voxel based game I played (Commanche) back in the DOS days and thinking how much potential it had if only they could get it to the point where the voxels weren't so darned big.

I wonder if some innovative/daring game designer will attempt a Cinema 2.0 (no polygon) based game.

If they have the budget for I'd think Adventure games would be perfect for this as they don't rely on speed. This is, of course, assuming Cinema 2.0 would be too slow for a FPS type of game, and possible even too slow for real time RPG or RTS.

Then again probably not unless Nvidia hardware can do it also.

Regards,
SB

 

[Jawed]

How did Guru3D get it working? Any ideas on what Hilbert did? I'm perma-banned from there so I can't ask and my buddy who wants to know says they've disabled registration.

http://www.guru3d.com/article/amd-ati-radeon-hd-4850-review-force-3d--powercolor/19

Hmm

Overly pessimistic devs?

I might have a go at playing the demo of Grid later, just downloaded it...

Jawed

 

[AlphaWolf]

How did Guru3D get it working? Any ideas on what Hilbert did? I'm perma-banned from there so I can't ask and my buddy who wants to know says they've disabled registration.

Here ya go

 

[mboeller]

High quality version is worthwhile

http://www.youtube.com/watch?v=ROAJMfeRGD4&fmt=18


Yeah he seemed to be saying they have a variety of maps, which are static.

Rendering, though, appears to be voxel-based ray-tracing with wavelets used to compress the voxels - running in real time. The whole thing is ray traced, there is no triangle-z-buffer based rendering. Seems pretty unbelievable...

Jawed

What about image based rendering like pixelplanes?

 

[AlexV]

If ever there's a game to test on HD4870X2, it would seem to be Racedriver GRID. From the FAQ:



http://community.codemasters.com/forum/showthread.php?t=279124

Jawed

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.

 

[Jawed]

What about image based rendering like pixelplanes?

I've got no idea. He says very little during the 3 minutes, so I'll guess we'll have to wait for the real launch next week.

Let's hope they don't hold this stuff back for X2' launch.

Jawed

 

[thatdude90210]

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

 

[SimBy]

If ever there's a game to test on HD4870X2, it would seem to be Racedriver GRID. From the FAQ:



http://community.codemasters.com/forum/showthread.php?t=279124

Jawed

It's suppose to scale great on CF:

"best optimized game for Crossfire i have ever played , not even 1 single break/stutering !

And look at performance @ 1920x1200 "

http://www.xtremesystems.org/forums/showthread.php?t=191574

 

[Jawed]

Note to SimBy (whose message got trashed). Yeah, seems those Grid developers really were living in oblivion to make that statement and then for everything to change. Bizarre.

I dunno it actually makes me wonder if some studios actively avoid dealing with the IHVs.

Jawed

 

[randomhack]

About the new hotfix for 4800 series: does anyone have information about whether its faster than the earlier version?