City Of The Future Tessellation demo

[blastingcap]

I had the same bad feeling that physx would be in Crysis 2 when I read about that $2 million NV paid Crytek to get their name on it.

Someone replied to me with some info about Crytek having already put a list of specifications of Cryengine 3 out there and it made specific mention of their own physics implementation.

I think it is a safe bet there will be no physx in Crysis 2. It will be the usual TWIMTBP logo on startup and of course the game will use DX11.

Crytek designs games so well I have my doubts they would ever add something to the detriment of their game, or moreso, their engine, just to satisfy NV. I'm sure there will be tessellation but I'll bet it will be to the benefit of the game rather than making an excessive use of it for the sake of marketing.

Good to know, I would look forward to it either way, even though I will doubtlessly get like 10fps on a single GTX460-768, haha.

 

[shangshang]

I can't believe that I just spent a good chunk of my time reading ALL the posts in here before me. All I gotta say is:

1) AMD software team has been trailing their counterparts over at NV... and it is of no surprise to me that NV is taking the thunder out AMD's tessellation. It should clearly be understood that when you are the "leader" in DX 11 (such as AMD), and tessellation is a major selling point of DX11,.. and you let some guy (like NV) who just stroll along and casually take your lunch and take over your turf (tessellation)... it surely has to reflect on the weak software team at AMD. I don't know how can anyone can under estimate this fact. People say let's wait until the 6900 series and then make the decision. Wait for the 6900?? I will bet NV will still kick AMD in the tessellation ass. AMD software team is not a miracle worker. Say what you want about NV practices, they have top notched software peeps, and to me NV just has grander visions on what a GPU should or should be capable of doing.

2) Scali seems to be the only guy in here who knows about tessellation well enough to expose AMD! I keep waiting for the next guy who could retort directly to Scali on a technical scale, but so far none in here could. Scali could very well be bs'ing too, but it would still take a person with knowledge as good as him to call him bs, and so far, no one has call him bs. So he must know something.

disclaimer: I ain't no graphics guru. In fact my profession is not even in computer. However, it's pretty easy to see AMD back peddling on the tessellation front! Lol and please don't ask me to "show my credential" before I can speaketh! Abstraction doesn't require such.

 

[blastingcap]

Here, read this: http://www.anandtech.com/show/2918/2

What Scali said was not anything new, but I understand why he's frustrated with having to explain the same thing over and over again, especially to people who never read the Fermi architecture articles and white paper.

AMD went for a kludgey fixed tessellator and no fundamental geometry engine change, but NV split up its geometry engine so that it would run more in parallel. Note that NV touts its geometry performance a lot (up to 8x faster than GT200 series), not just its tessellation performance. Tessellation is part of a greater whole but gets talked about a lot by the pro-NV crowd because AMD's geometry engine is fine for current games but fails in extreme tessellation.

Why did AMD do this? I suspect because they were on a tight timeline (they wanted Cypress launched no later than Windows 7's launch). Making their own polymorph engine probably took so much time that they gave up and used a kludgey band-aid (fixed tessellator). Now that AMD is no longer on a tight timeline and is getting hammered by NV in tessellation PR, I have no doubt that they are working hard to implement their own parallel or quasi-parallel geometry engine.

I can't believe that I just spent a good chunk of my time reading ALL the posts in here before me. All I gotta say is:

1) AMD software team has been trailing their counterparts over at NV... and it is of no surprise to me that NV is taking the thunder out AMD's tessellation. It should clearly be understood that when you are the "leader" in DX 11 (such as AMD), and tessellation is a major selling point of DX11,.. and you let some guy (like NV) who just stroll along and casually take your lunch and take over your turf (tessellation)... it surely has to reflect on the weak software team at AMD. I don't know how can anyone can under estimate this fact. People say let's wait until the 6900 series and then make the decision. Wait for the 6900?? I will bet NV will still kick AMD in the tessellation ass. AMD software team is not a miracle worker. Say what you want about NV practices, they have top notched software peeps, and to me NV just has grander visions on what a GPU should or should be capable of doing.

2) Scali seems to be the only guy in here who knows about tessellation well enough to expose AMD! I keep waiting for the next guy who could retort directly to Scali on a technical scale, but so far none in here could. Scali could very well be bs'ing too, but it would still take a person with knowledge as good as him to call him bs, and so far, no one has call him bs. So he must know something.

disclaimer: I ain't no graphics guru. In fact my profession is not even in computer. However, it's pretty easy to see AMD back peddling on the tessellation front! Lol and please don't ask me to "show my credential" before I can speaketh! Abstraction doesn't require such.

 

[cbn]

Here is a tessellation demo that nVidia has shown at GTC 2010 recently:
http://www.youtube.com/watch?v=Vz1Y5YkqgyQ

It should give you a good impression of what tessellation can really do, when the whole engine and geometry is designed for it (unlike many current titles, which are basically just console-ports with fake tessellation bolted on).
Note how shockingly low the detail is when the tessellation is disabled. All that detail is generated by the GPU on-the-fly, rather than being stored in memory.

I thought the quote of 80 GB/s to render the tessellation of the first demo sounded super-impressive. Of course, JHH is saying this is well beyond that what PCI-Express has to offer.

What kind of interface then does Nvidia plan to run these kind of programs on?

 

[Lonbjerg]

I thought the quote of 80 GB/s to render the tessellation of the first demo sounded super-impressive. Of course, JHH is saying this is well beyond that what PCI-Express has to offer.

What kind of interface then does Nvidia plan to run these kind of programs on?

The GPU...you should read the thread, it's all explained in detail?


And what interface do you think they used for the Fermi GPU's running the demo...a new, unknown newly invented NVIDIA exclusive interface?

 

[PingviN]

1) AMD software team has been trailing their counterparts over at NV...

AMD's tesselation is not bottlenecked by poor software. It's bottlenecked by hardware.

 

[blastingcap]

I thought the quote of 80 GB/s to render the tessellation of the first demo sounded super-impressive. Of course, JHH is saying this is well beyond that what PCI-Express has to offer.

What kind of interface then does Nvidia plan to run these kind of programs on?

Oh come on, man, did you even read Scali's explanations in this thread? No wonder he gets frustrated.

 

[blastingcap]

AMD's tesselation is not bottlenecked by poor software. It's bottlenecked by hardware.

They are interrelated. The tough part is to get multiple geometry engines to talk to each other and spit out a coherent result.

Since AMD's CPU team has experience with out-of-order execution, I wonder if the GPU team could "borrow" some of the CPU staff to make a truly parallel geometry engine and thus outdo NV's Polymorph engine? Hmm.

 

[Dark Shroud]

I'm waiting for 6900 series before I judge the performance of the 6000 series in general, especially in tessellation. The fact that the 6800 does series have updated tessellation performance shows that AMD is at least thinking about this.

I expect by AMD's 28nm GPU launch they will have their own truly parallel geometry engine to match or exceed Nvidia's engine.

With the performance updates that AMD & Nvidia have been making 2GB standard memory on highend models & PCI-Express 3 can't get here soon enough.

Lastly I'll go out on a limb here and say that Crysis 2 will probably have lots of tessellation going all the way up to extreme. Only they'll use a slider option for current limited cards to future cards.

 

[Scali]

As Scali has rightfully pointed out, though, AMD's kludgey solution is not as elegant or "futureproof" as NV's.

Thank you. At least someone acknowledges the point I'm trying to make.

 

[Scali]

Good to know, I would look forward to it either way, even though I will doubtlessly get like 10fps on a single GTX460-768, haha.

You know what?
I wouldn't even mind!
I replayed the original Crysis at least four times:
1) At release, on my 8800GTS in DX9 mode.
2) After a few updates of Vista, the game itself, and the drivers, when DX10 finally became a feasible option.
3) With my Radeon 5770 now that I could run with 4xAA, better framerates, and even better image quality.
4) With my GeForce GTX460 and my new 1080p monitor.

 

[blastingcap]

Thank you. At least someone acknowledges the point I'm trying to make.

I don't doubt your technical knowledge, I just wish you wouldn't be as abrasive sometimes. (A mod made me take your quote out of my signature block.) And of course I appreciate your point; I read the Fermi analysis at AT a long time ago, so I knew what you were talking about re: parallel Polymorph engines. That's why I am here and not at, say, HardOCP, which is very good at what it does but hardly ever gets really in-depth re: things like architecture and hardware engineering.

As for Crysis 2, I think I'm going to wait till I have Kepler or a HD7xxx or something to even try to play that, if it's going to murder hardware like Crysis did.

 

[Scali]

AMD went for a kludgey fixed tessellator and no fundamental geometry engine change, but NV split up its geometry engine so that it would run more in parallel. Note that NV touts its geometry performance a lot (up to 8x faster than GT200 series), not just its tessellation performance. Tessellation is part of a greater whole but gets talked about a lot by the pro-NV crowd because AMD's geometry engine is fine for current games but fails in extreme tessellation.

It's a bit of a catch-22 problem.
If you don't have tessellation, you need to feed every triangle directly from memory.
You have neither the storage nor the memory throughput to handle such large triangle-counts that way, so geometry engine bottlenecks would not show up before memory bottlenecks.

But as you want to push for more geometry, tessellation is really the only way.
I mean, I've mentioned Pixar/RenderMan a few times already. For them it is EVEN MORE impossible to store their geometry in memory. They have less than one triangle for every pixel, and that is with a lot of AA, and at resolutions far higher than what our simple computer screens are capable of.
It would be completely impossible to manage the geometry directly. You couldn't store it, couldn't model it, couldn't animate it, nothing. You would need terabytes and terabytes of memory to store it.
But thanks to tessellation, they can refine their geometry detail to perfection, literally.

 

[Scali]

I just wish you wouldn't be as abrasive sometimes.

The constant slurs and denial made me lose my patience with a number of people a long time ago.
I am not being treated with any kind of respect at all. People can expect to be met with the same amount of respect that they grant me.
Let's face it, a number of people here are only defending 'their brand'. They don't know why, because they don't understand the technical background. Therefore they also cannot argue any technical points I'm making. So they just throw personal insults, trolls and thread derailments around, all the time. I hardly ever get anyone to back up what I say, no matter how right I am, and no matter how often I point to authoritative sources, such as Microsoft themselves, saying exactly what I am saying.
I wonder how you'd be in a similar situation.

(A mod made me take your quote out of my signature block.)

Seems like you can be abrasive as well.

 

[Lonbjerg]

The constant slurs and denial made me lose my patience with a number of people a long time ago.
I am not being treated with any kind of respect at all. People can expect to be met with the same amount of respect that they grant me.
Let's face it, a number of people here are only defending 'their brand'. They don't know why, because they don't understand the technical background. Therefore they also cannot argue any technical points I'm making. So they just throw personal insults, trolls and thread derailments around, all the time. I hardly ever get anyone to back up what I say, no matter how right I am, and no matter how often I point to authoritative sources, such as Microsoft themselves, saying exactly what I am saying.
I wonder how you'd be in a similar situation.

I have seen you post elsewhere and have always liked your insigths and I wouldn't worry to much though.
The people attacking you and not your arguments are the people who have no valid counter-arguments.

I am sure that most of the silent readers of the thread will agree to that part, you don't need a PhD to figure that one out

 

[cbn]

Oh come on, man, did you even read Scali's explanations in this thread? No wonder he gets frustrated.

Sorry, I didn't have time yet.

 

[flopper]

They are interrelated. The tough part is to get multiple geometry engines to talk to each other and spit out a coherent result.

Since AMD's CPU team has experience with out-of-order execution, I wonder if the GPU team could "borrow" some of the CPU staff to make a truly parallel geometry engine and thus outdo NV's Polymorph engine? Hmm.

Likely things be happening at 28nm for that pararell stuff to happen more.
6970 wont likely have that even if its improved.
AMD are releasing things on schedule, which dosnt allow them much headroom to add a highly technical feature but when they do we will hear about it

 

[cbn]

The GPU...you should read the thread, it's all explained in detail?


And what interface do you think they used for the Fermi GPU's running the demo...a new, unknown newly invented NVIDIA exclusive interface?

Okay, I *think* got it now.

The GPU is reducing the need for PCI-Express bandwidth through tessellation.

If tessellation were not present achieving the same level of quality would have far exceeded the capability of the PCI-E.

 

[AtenRa]

A game that uses tessellation to its fullest extent would have low poly meshes though, if you ported it to a system that can't tessellate using the displacement maps you would need to design additional high quality models or continue with what we've been seeing so far - high quality models with tessellation bolted on to reach very high quality.

Games already have multiple levels of poly count, they use different poly count for adaptive transformation. In Civ-5 the game can be played in DX-9 or DX-10 mode but in DX-11 mode will be faster because of tessellation.
Civ-5 could be ported to Consoles because it’s already have high poly mess in order to be backwards compatible with DX-9 and DX-10 hardware.

So we can have a DX-11 low poly game and at the same time have high poly meshes for backwards compatibility and for Console porting.

 

[shangshang]

excuse me this may be a dumb quesiton, but I'm understanding that through tessellation usage, a game developer does not need to make as many models, freeing himself some time, because he can just let the GPU do its thing, i.e, tessellate! I wonder if this will translate into lower production cost for developer, and higher quality "game play" (because the freed up time could now be put to creating a great storyline?).

 

[Scali]

excuse me this may be a dumb quesiton, but I'm understanding that through tessellation usage, a game developer does not need to make as many models, freeing himself some time, because he can just let the GPU do its thing, i.e, tessellate! I wonder if this will translate into lower production cost for developer, and higher quality "game play" (because the freed up time could now be put to creating a great storyline?).

I don't think it would make that much of a difference, as tessellation has been part of modeling software for ages, so generating geometry at multiple details could already be automated.
It's now easier to 'fine-tune' though, as you can modify the detail in realtime on the GPU. Previously you had to precalc various levels of detail.

 

[blastingcap]

Endless city hack now allows Radeons to run the demo:

http://scalibq.wordpress.com/2010/11/25/running-nvidias-endless-city-tessellation-demo-on-radeons/

 

[VirtualLarry]

Endless city hack now allows Radeons to run the demo:

http://scalibq.wordpress.com/2010/11/25/running-nvidias-endless-city-tessellation-demo-on-radeons/

Please start a new thread on that, I'm afraid this post will get lost in this thread.

Edit: I wasn't able to get it to work. The NV installer automagically installs the 64-bit EXE onto my Win7 64-bit OS, and the patch only works with the 32-bit version.

 

[Leadbox]

The artwork does not impress me, not the best way to showcase the nvdia's tessellation implementation. Why didn't they redo Unigene, tessellate the entirety of it that would have a real wow for me

 

[taltamir]

Please start a new thread on that, I'm afraid this post will get lost in this thread.

I second this. this is a really awesome patch and it would great to compare the framerates of different hardware from nvidia and amd while running the endless city when patched to be vendor agnostic.