AMD: R9xx Speculation

Why do you think it? If these results reflect tesselation performance, it should be pretty close to HD5700.

 

I just took the 20 fps gap between the 5870 and 5770 and assumed it would be similar on a 5670 compared to the 5770. Perhaps saying unplayable wasn't exactly true, but you sure wouldn't be happy with it.

 

It probably wouldn't run all that well on a GT 430 either…

 

I'm surprised at huge drop in FPS for nV at higher AA. At 8xAA GTX 460 is barely faster than HD 6870.

 

jimbo75: HD5800 has dual rasterizer, which may somewhat increase tessellation efficiency. But HD5700 and 5600 both have one rasterizer - if the performance is really limited by GPU's front-end (tesselation/setup/rasterizer), their framerates shouldn't be much different.

 

Yes true enough, but the 5670 is also in-between clocks at 775mhz.

Look at the gap between the 5870 and 5850 (big), then the gap between the 5850 and 5830 (small). Looks more like clock speed is a limit, but you can never tell with benchmarks.



http://img.hexus.net/v2/graphics_cards/nvidia/Fermi/a/N174/Mod.jpg

http://img.hexus.net/v2/graphics_cards/nvidia/Fermi/a/N174/Norm.jpg

http://img.hexus.net/v2/graphics_cards/nvidia/Fermi/a/N174/Xtreme.jpg

If you put each of those in a tab you can see the difference - and that's at extreme in a tessellation benchmark. I just doubt it at all that anybody would notice the difference during gaming in Hawx 2.

If AMD can reduce the tesselation levels in drivers, show how little the IQ loss is, they will surely be able to hugely increase fps over Fermi in Hawx 2.

 

Gt 430 isn't a gaming card and never will be. The 5670 is, for a lot of people I'd guess.

 

@Simby

Well if there is a noticable IQ loss I suppose it's fair enough. We'll see on that.

For me, this is why Nvidia rounded on the FP16/FP11 thing at the 450's release. Talk of ATI's "cheat" and some sites testing Catalyst AI - this is linked to it too.

Nvidia is trying to make tiny IQ differences which result in large fps gains seem like "cheating", when in actual fact anybody with common sense would take the "cheat" fps compared to the negligible IQ difference.

 

You've gone back to square one. Read this discussion again and then get back to me.

Indeed, even more reason for them to avoid talking about nVidia's tessellation advantage (real or imagined).

If the original is anything to go by that advantage is not just due to tessellation. The first one liked nVidia's hardware too.

 

H.A.W.X. is heavy on geometry.

 

I wonder if it's possible to simply discard everything above what AMD thinks is not excessive, on the driver level. I'm sure everyone would be yelling cheating but if there's not much difference in IQ it should prove AMDs point.

I have a feeling it's not possible though.

 

Couldn't they just limit the max tesselation factor to desired at driver level on .exe basis? (so everything asking higher tesselation factor would just get the max set level instead of what DX API allows)

 

The dual rasterizer makes no difference in tessellation performance. No test shows any Evergreen chip generating more than one tessellated triangle
every three clocks.

I'm pretty sure that the dual rasterizer is there for easier data transport across the chip. Having a single rasterizer and transferring pixel data is probably more costly.

 

Even after AMD somehow patches HAWX 2 it seems that a 6850 could still beat 5870 .
If future tesselated games will reflect the order of AMD cards in the actual benchmark than 6850 is faster card tha 5870 and the new naming scheme fits quite good.

 

How are these results high for AMD cards? For me it's looks like the difference between AMD and nVidia cards in Hawx 2 can not be explained by pure tesselation power.

See this for comparison

Or is Heaven not representative at all for tesselation capacity?

 

Er, actually, I'm not. I'm not in agreement with AMD's stance till I see some evidence. I think tessellation in Evergreen is a kludge and they better fix it in Cayman.

Isn't that a Microsoft sample, part of the SDK? Like the other D3D11 samples shown before Evergreen's launch?

Apparently they're doing it, with games like Civ 5.

The era of demos for tessellation is over. By several years.

Actually, what AMD's apparently accusing NVidia of doing is unnecessarily hobbling performance for the 80%+ of D3D11 gamers out there. But NVidia hobbling performance and IQ for people who aren't using NVidia is nothing new.

From the presentation that Sontin linked there's no doubt that there's too much tessellation in certain areas of the screen, but the algorithm is clearly much better than the naive rubbish seen in Heaven. I'm not convinced that HAWX2's intrinsically bad (it's a delicate balance and very hard to avoid over- and under-tessellation simultaneously), so I'm waiting to see why AMD thinks this game is excessive, quantitatively and qualitatively. There are more advanced techniques, such as making the silhouette the focus of the highest-quality tessellation. In fact NVidia went into this technique in some detail over a year ago. So you have to ask why that isn't in the game.

I also think that Huddy's talking nonsense about the required size of triangles. That's just an excuse, those are triangle sizes that don't disappear. I'm not convinced AMD has an argument. Some would argue that "good-enough" is OK, no need to stress out the hardware, but GTS450 seems to be OK.

The stuff about fragment over-shading is very true, but that's because quad-based rendering is too granular (it's a similar problem to dynamic branching incoherence). I suspect ATI's architecture is especially inefficient in the management of quads in hardware threads, so adding to the pain.

Tessellation performance of Evergreen might have been adequate back in R600's days, but a forward-looking architecture's required. It doesn't scale.

The "price" NVidia's paying for advanced tessellation/setup/rasterisation is not dissimilar to the price AMD paid with out-of-order thread execution in R520 and the fine-grained dynamic branching. That price has to be paid, going forwards.

Tessellation is about rendering the right-sized triangles. They seem right-sized to me in HAWX2, but that needs careful evaluation - and the algorithm may be on the naive side. If the game has a tessellation-quality slider then gamers can choose their trade-off and we can argue about IQ

 

Charlie is not happy and critisizes AMD.



AMD breaks it's own naming scheme with the 6870
Good chip, bad naming

by Charlie Demerjian

 

well amd makes more money selling it as a 6800 serie than a 6700 serie, so I woulndt think it was a hard decision to rename the cards.

wait a few weeks, and then the 6970 have great performance, and not many would care at all.

 

Actually, AMD is accusing NVidia of hobbling performance for 100% of the D3D11 gamers out there. NVidia's cards are also running slower than they have to, though it just makes less of a difference to them.

Looking at the dev talk that Sontin pointed to, the tessellation really is quite excessive. There's already too many triangles for the mountains, and they say that they cranked down the tessellation for the audience to see the wireframe. Adaptive tessellation is used, but there's no adaptation for patches that have no chance of generating silhouette triangles (and should thus use non-occlusion parallax mapping).

Still, I do want to see AMD get a higher rate of setup. Cayman will hopefully do the job if that leaked line of "scalability" holds true. Just because HAWX2 has unnecessarily excessive triangle count doesn't mean there aren't other legit 10M-triangle scenes in games.

At the very least, AMD should be able to cull/clip more than one tri per clock. The hardware cost and complexity of that is trivial, as there's no ordering issues to worry about.

 

Although AMD could have done better naming choice, but honestly most of us probably wont be able to come up with better choice considering:

1. There are more GPUs to fit into naming schemes, including Fusion taking the bottom. From this perspective AMD did a right choice, even though it leaves little room for Antilles.

2. Naming schemes should be favorably accepted by buying customers, and 68xx definitely would sell better than 67xx, thats the sleazy part. Still AMD havent made the biggest sin - asking higher prices for Barts, they could have easily asked ~GTX470 price for 6870 and got away with it, that would have been wrong.

In the end, we got very nice upgrade for 5700 gen, and forced Nvidia to make significant price cuts, win-win for customers. Outside of minor naming controversy that was a great launch.

P.S. Charlie keeps mentioning as if 6800 is successor for 5800, its not. Also he demands Barts to be priced lower