NVIDIA GF100 & Friends speculation

[Silus]

I doubt any game in next year or so will use tesselation heavily enough to give significant benefits for GF100. More likely it'll be like how X1900's beat the 7900's nowadays, but it's well past their age.

I sort of agree. While Microsoft doesn't roll out a new console with a graphics chip that supports DX11, games that use tessellation won't be many, though I wouldn't go as far as saying that in a year we won't even see one game that uses it

NVIDIA's relations with developers, may play an important part as well and we may see more games that use tessellation.

 

[CRoland]

Two ? Cypress only has one.

Is this wrong?
http://www.geeks3d.com/20100210/tes...n-opengl-radeon-hd-5000-tessellators-details/

On Radeon HD 5000 series (Evergreen family), things are different. The Radeon HD 5000 includes the fixed tessellator of HD 2000, 3000 and 4000 AND a new programmable tessellation unit. This new programmable tessellation unit will be exposed via a new OpenGL extension (available shortly).

 

[Kaarlisk]

More likely it'll be like how X1900's beat the 7900's nowadays, but it's well past their age.

I have never managed to find benchmarks that show this. Could someone please point them out?

Also, the Radeon X1950pro, which was a mainstream card at the time, is still used by quite a few people I know - so they are now enjoying the benefits of such forward thinking.

 

[Arty]

If there's any truth in that graph then it's 50-60% faster than a 285. Not great, but not a disaster either. Kinda boring after the long wait but I'll reserve final judgment until I see how it fares in the titles where it matters.

Anything less than a 30% performance delta with Cypress will be a disaster.

 

[mboeller]

I have never managed to find benchmarks that show this. Could someone please point them out?

Also, the Radeon X1950pro, which was a mainstream card at the time, is still used by quite a few people I know - so they are now enjoying the benefits of such forward thinking.

In German language:
http://www.computerbase.de/artikel/hardware/grafikkarten/2009/test_14_grafikkarten_vergleich/

 

[Silus]

Is this wrong?
http://www.geeks3d.com/20100210/tes...n-opengl-radeon-hd-5000-tessellators-details/

Never heard of that. There's no mention here:

http://www.beyond3d.com/images/reviews/cypress-arch/cypress-arch-big.png

Tech-Report's GPU and Architecture analysis includes the pics from AMD itself:

http://techreport.com/r.x/radeon-hd-5870/block-diagram.jpg

There's also no mention of this additional tessellation unit. All both say is that the fixed tessellator unit is quite a bit different than the one found since R600.

 

[trinibwoy]

Not that Cypress had perfect scaling in that sense from rv770, but it certainly doesn't look like nvidia could make that perf/transistor deficit they have since rv7xx any smaller this generation.

Yeah if it manages to hit 50% more than a 285 it'll be in the same ballpark as Cypress' (in)efficiency relative to RV770.

Yeah, definitely not great if true across the board. But it's not that bad either. I do want to see the tesselation results in actual games though, since that is apparently where it should show its muscle.

Well I would begin to care if developers were actually doing something useful with tessellation. Dirt2, Stalker, AvP and now Metro have all utterly failed in that regard IMO.

Anything less than a 30% performance delta with Cypress will be a disaster.

30% over Cypress is nipping on Hemlock's heels. That's far from disaster territory. Even if it's only 10% faster they will sell every one they can make though that is arguably not a difficult proposition given the continued doom and gloom over Fermi yields.

 

[PatrickL]

Anything less than a 30% performance delta with Cypress will be a disaster.

What if Nv does not care to have the best graphic card but fast just enough while promoting the gpgu part ? If they think that gpgpu computing will be their future market, they could try the same strategy Sony made with the PS3 to win the Bluray war

 

[Malo]

I sort of agree. While Microsoft doesn't roll out a new console with a graphics chip that supports DX11, games that use tessellation won't be many, though I wouldn't go as far as saying that in a year we won't even see one game that uses it

NVIDIA's relations with developers, may play an important part as well and we may see more games that use tessellation.

I'd have to disagree, whilst the usage hasn't been terribly complete, from my pov it looks as though there are more games with tessellation and/or direct compute, or even things like DX11 shadows being used to some degree already compared to DX10 usage when it was first released. And we only have ATI DX11 cards. Imagine when Nv finally releases Fermi?

 

[Malo]

30% over Cypress is nipping on Hemlock's heels. That's far from disaster territory. Even if it's only 10% faster they will sell every one they can make though that is arguably not a difficult proposition given the continued doom and gloom over Fermi yields.

10% would definitely be a disaster, because they would have to price is accordingly which would kill their profits given the chips size wouldn't it?

 

[Silus]

Well I would begin to care if developers were actually doing something useful with tessellation. Dirt2, Stalker, AvP and now Metro have all utterly failed in that regard IMO.

Call of Prypiat also used Tessellation ? Don't think I saw the results on that one then. Dirt 2 no doubt I agree. Nothing impressive. AvP was better though also nothing impressive.

Still, obviously the need for its use by developers is the key« and that's why NVIDIA's TWIMTBP program may play a relevant role. I don't think NVIDIA made GF100 such a geometry crunching monster on paper and won't even try to show it in games.

 

[trinibwoy]

10% would definitely be a disaster, because they would have to price is accordingly which would kill their profits given the chips size wouldn't it?

After following these proclamations for a while I've come to the solid conclusion that nobody knows anything when it comes to IHV costs. It's usually a bunch of huffing and puffing that never materializes in financial results. For example, where is all the money that AMD was supposed to make on those cheap dies last generation? Where are those fantastic margins?

Besides, Nvidia can safely sell a card that is 10% faster for 20% more money. They've got the name, reputation, marketing and "single-fastest GPU" halo to hold on to.

Still, obviously the need for its use by developers is the key« and that's why NVIDIA's TWIMTBP program may play a relevant role. I don't think NVIDIA made GF100 such a geometry crunching monster on paper and won't even try to show it in games.

Hence my confusion at Metro. The GF100 whitepaper touted displacement mapping as the second coming and yet it's nowhere in sight. Just ramping up the polycount to get smoother objects is lame since that isn't noticeable without static screenshots. And even then it's hard to see.

 

[Silus]

I'd have to disagree, whilst the usage hasn't been terribly complete, from my pov it looks as though there are more games with tessellation and/or direct compute, or even things like DX11 shadows being used to some degree already compared to DX10 usage when it was first released. And we only have ATI DX11 cards. Imagine when Nv finally releases Fermi?

DX10 had other problems, like not being able to be used under XP, which crippled its "adoption" rate.
Although I do think that DX11 will be used more often than DX10, it won't be that big of a difference and I tie that fact with consoles. The lowest common denominator will play its part on the gaming market once again.

 

[Jawed]

Yeah if it manages to hit 50% more than a 285 it'll be in the same ballpark as Cypress' (in)efficiency relative to RV770.

HD5870 ~90%+ faster than HD4890 not enough for you?

http://www.pcgameshardware.de/aid,7...ate-GPU-PhysX-Video/Action-Spiel/Test/?page=2

Jawed

 

[mboeller]

5970,5870,5870 2GB 5850 VS. 480, 470 final scores:

Hi;

do you know what resolution they used for the tests? I think the test was done @ 1920x1200 because the difference between the 2GB and 1GB HD5870 is rather small. Also, was AA used for the tests (the HD5830 is much slower than the HD5850 and that happens mostly due to AA AFAIK)?

 

[trinibwoy]

HD5870 ~90%+ faster than HD4890 not enough for you?

http://www.pcgameshardware.de/aid,7...ate-GPU-PhysX-Video/Action-Spiel/Test/?page=2

I thought it was clear we were talking about averages? Who's to say GF100 won't be 120% faster than GT200 in some game (might even be Metro)?

 

[mczak]

There's also no mention of this additional tessellation unit. All both say is that the fixed tessellator unit is quite a bit different than the one found since R600.

I think there's just some misunderstanding somewhere. Basically, the fixed part of the tesselation unit is basically the same, whereas the programmable part is new.

I'm sure that that will confirmed, with the newer games that should use heavy tesselation.

It certainly should excel with tesselation, though I'm unsure how much better it will really be with heavy tesselation. Fermi certainly appears to be faster in those unigine scores, but it's nowhere near 3x which the theoretical tri throughput would suggest.
Also, I wouldn't expect this to be a problem for owners of HD5xxx. Maybe some games will have configurable tesselation, with the highest level really taxing those cards, but possibly with quite low image quality impact (diminishing returns the higher the tesselation level is).

Maybe the problem is the missing TMUs, according to Neliz?

That's too mysterious to really comment on it. If it really is missing texturing power (the tex:alu ratio is quite similar to Evergreen family FWIW), maybe derivatives not suffering from this could have higher perf/transistor. But then Juniper is also quite a bit better than Cypress there, though of course we don't really know if Fermi doesn't also suffer from mysterious scaling problems like lack of internal bandwidth (though on paper it doesn't look so).

 

[Jawed]

I sort of agree. While Microsoft doesn't roll out a new console with a graphics chip that supports DX11, games that use tessellation won't be many, though I wouldn't go as far as saying that in a year we won't even see one game that uses it

XB360 has a tessellator. Let's see if it'll get some use (other than in about 2 games if I recall right), now that D3D11 has arrived...

Jawed

 

[Arty]

30% over Cypress is nipping on Hemlock's heels. That's far from disaster territory. Even if it's only 10% faster they will sell every one they can make though that is arguably not a difficult proposition given the continued doom and gloom over Fermi yields.

I'm not talking from a sales or financial pov but from an overall performance pov. A smaller delta would mean Nvidia would be in a worse position than they were in with the GT200. Being 50% larger than Cypress, my expectations perfectly reasonable. I wont even get Hemlock into the discussion as thats not even relevant here.

 

[Jawed]

I thought it was clear we were talking about averages?

Why use games that scale badly in your averages?

Jawed