Is there any truth that AMD cards are bad at Tessellation

Kaapstad · 15 Mar 2014 at 22:18

Well lets find out

What would be helpful is if some people can run the Heaven 4 bench with everything maxed except the tessellation which should be set to off. If anyone has a GTX 780, GTX 780ti or Titan and can do this, when i get home from work in the morning I will do the same with my 290X.

In theory if the AMD cards are bad at tessellation, by turning it off in the heaven 4 bench then my 290X should be able to close the gap.

If anyone is up for it I think this will be an interesting exercise and may answer a few questions.

LtMatt · 15 Mar 2014 at 22:19

Good idea Kapston. I'll be up for this... when i get a working motherboard.

Kaapstad · 15 Mar 2014 at 22:21

LtMatt said:
Good idea Kapston. I'll be up for this... when i get a working motherboard.

Cheers Lt.

shankly1985 · 15 Mar 2014 at 22:21

LtMatt said:
Good idea Kapston. I'll be up for this... when i get a working motherboard.

What's wrong with your motherboard?

drunkenmaster · 15 Mar 2014 at 22:26

Kaapstad said:
Well lets find out

What would be helpful is if some people can run the Heaven 4 bench with everything maxed except the tessellation which should be set to off. If anyone has a GTX 780, GTX 780ti or Titan and can do this, when i get home from work in the morning I will do the same with my 290X.

In theory if the AMD cards are bad at tessellation, by turning it off in the heaven 4 bench then my 290X should be able to close the gap.

If anyone is up for it I think this will be an interesting exercise and may answer a few questions.

No there isn't, there is however Nvidia and there general shady stupidity of ruining everything for the sake of winning.

Tessellation is predictable and provable. Grab a piece of paper, draw a big triangle, split it into two more, keep doing this. You see how eventually you get to a point that you can't actually see new triangles? What happens is, there absolutely, categorically IS a limit beyond which tessellation is absolutely worthless. So sensible hardware makers design a tessellation unit to work effectively at this level and waste no hardware on levels you can't ever see......

Then you have Nvidia, why not over engineer the tessellator, then pay a few games to use tessellation WAY beyond a level you can physically see... but your hardware will cope great with it and AMD hardware won't.

It's silly games, it's not realistic, you can only induce a massive performance delta between AMD/Nvidia via tessellation at a level you can see zero effect from.

Benchmarking is pointless, enable the tessellation limit in control panel and AMD drivers should limit the excess and useless tessellation from being done, I don't know if it would effect benchmark results. But it's now Nvidia are getting such good tessellation results. Personally I don't care how fast my card can tessellate beyond a level of detail no screen made today can actually show.

This is exactly no different to what Nvidia did with Crysis 2(I forget which one it was to be honest, 2 or 3 I think anyway). Massively massively over tessellate a completely flat object to make something that topples AMD hardware, doesn't topple the Nvidia tessellator, but offered precisely zero IQ improvement. We're talking about Nvidia pushing tessellation up to sub pixel changes that can't possibly be seen... the fact that Nvidia is wasting transistors to do this faster just to be able to mess around in benchmarks is a joke.

There is a crap way to do tessellation, a smart way, and a completely mental only out to cheat in benchmarks way... guess which way Nvidia went.

Orangey · 15 Mar 2014 at 22:33

Does Heaven 4 have this sub-pixel tesselation?

Kaapstad · 15 Mar 2014 at 22:34

drunkenmaster said:
No there isn't, there is however Nvidia and there general shady stupidity of ruining everything for the sake of winning.

Tessellation is predictable and provable. Grab a piece of paper, draw a big triangle, split it into two more, keep doing this. You see how eventually you get to a point that you can't actually see new triangles? What happens is, there absolutely, categorically IS a limit beyond which tessellation is absolutely worthless. So sensible hardware makers design a tessellation unit to work effectively at this level and waste no hardware on levels you can't ever see......

Then you have Nvidia, why not over engineer the tessellator, then pay a few games to use tessellation WAY beyond a level you can physically see... but your hardware will cope great with it and AMD hardware won't.

It's silly games, it's not realistic, you can only induce a massive performance delta between AMD/Nvidia via tessellation at a level you can see zero effect from.

Benchmarking is pointless, enable the tessellation limit in control panel and AMD drivers should limit the excess and useless tessellation from being done, I don't know if it would effect benchmark results. But it's now Nvidia are getting such good tessellation results. Personally I don't care how fast my card can tessellate beyond a level of detail no screen made today can actually show.

This is exactly no different to what Nvidia did with Crysis 2(I forget which one it was to be honest, 2 or 3 I think anyway). Massively massively over tessellate a completely flat object to make something that topples AMD hardware, doesn't topple the Nvidia tessellator, but offered precisely zero IQ improvement. We're talking about Nvidia pushing tessellation up to sub pixel changes that can't possibly be seen... the fact that Nvidia is wasting transistors to do this faster just to be able to mess around in benchmarks is a joke.

There is a crap way to do tessellation, a smart way, and a completely mental only out to cheat in benchmarks way... guess which way Nvidia went.

I am not sure that AMD cards are as bad at Tessellation as some people say, I find my 290Xs seem to do quite well with Tessellation from the limited amount i have seen. This is the reason for doing the above exercise to see if it is the Tessellation in Heaven holding the cards back or it is something else.

Defy Belief · 15 Mar 2014 at 22:53

I'll do this.

Wouldn't it make sense to do a bench without tessellation then one with? And work out the percentage it impacts the card having it turned on?

Orangey · 15 Mar 2014 at 22:54

13.11 sig rig all stock, xfire disabled, settings as advised

Sugarhell · 15 Mar 2014 at 23:07

drunkenmaster said:
No there isn't, there is however Nvidia and there general shady stupidity of ruining everything for the sake of winning.

Tessellation is predictable and provable. Grab a piece of paper, draw a big triangle, split it into two more, keep doing this. You see how eventually you get to a point that you can't actually see new triangles? What happens is, there absolutely, categorically IS a limit beyond which tessellation is absolutely worthless. So sensible hardware makers design a tessellation unit to work effectively at this level and waste no hardware on levels you can't ever see......

Then you have Nvidia, why not over engineer the tessellator, then pay a few games to use tessellation WAY beyond a level you can physically see... but your hardware will cope great with it and AMD hardware won't.

It's silly games, it's not realistic, you can only induce a massive performance delta between AMD/Nvidia via tessellation at a level you can see zero effect from.

Benchmarking is pointless, enable the tessellation limit in control panel and AMD drivers should limit the excess and useless tessellation from being done, I don't know if it would effect benchmark results. But it's now Nvidia are getting such good tessellation results. Personally I don't care how fast my card can tessellate beyond a level of detail no screen made today can actually show.

This is exactly no different to what Nvidia did with Crysis 2(I forget which one it was to be honest, 2 or 3 I think anyway). Massively massively over tessellate a completely flat object to make something that topples AMD hardware, doesn't topple the Nvidia tessellator, but offered precisely zero IQ improvement. We're talking about Nvidia pushing tessellation up to sub pixel changes that can't possibly be seen... the fact that Nvidia is wasting transistors to do this faster just to be able to mess around in benchmarks is a joke.

There is a crap way to do tessellation, a smart way, and a completely mental only out to cheat in benchmarks way... guess which way Nvidia went.

Great post. I remember that crytek for ryse son of rome they reduce the triangles from 150k to 85k because the iq was the same. And they gained better shading + no LODs.Triangles have a diminishing return when you go over 50k. Its not like 15 years ago that double triangles (60 triangles to 120) resulted a way better mesh.

Defy Belief · 15 Mar 2014 at 23:20

http://postimage.org/app.php

Lightning 780 @ stock.

Just out of interest to see with it at extreme:

http://postimage.org/

Quite a hit but the score looks low in comparison to what I've seen most 780s do, need to check the BIOS not sure what BIOS 8 pack put on it before shipping.

humbug · 15 Mar 2014 at 23:21

I don't have a 290X but i'm still up for it, for whatever its worth.

From the Heaven 4 thread: Score 796, GPU 7870XT @1215/1600

Without Tessellation: 1200 / 1600 Score = 1061 (+33%)

I would be interested to see what level of performance gain a GTX 660TI has from turning off Tessellation.

Rroff · 15 Mar 2014 at 23:22

GTX780 (not at stock). Pretty sure I can get that upto atleast 90 as I wasn't quite at max clocks - 1250ish on the core.

arc@css · 15 Mar 2014 at 23:29

I think it was just mainly the 5000 series that were a bit pants at it.

Orangey · 15 Mar 2014 at 23:37

Sugarhell said:
Great post. I remember that crytek for ryse son of rome they reduce the triangles from 150k to 85k because the iq was the same. And they gained better shading + no LODs.Triangles have a diminishing return when you go over 50k. Its not like 15 years ago that double triangles (60 triangles to 120) resulted a way better mesh.

*THAT image*

That image has been debunked, if your starting mesh was designed with 60 triangles in mind then 60,000 means you start over and not try and gild a turd. I think Nvidia do have a leg to stand on as far as pushing boundaries but not with tesselated water hidden under the map out of sight.

whyscotty · 15 Mar 2014 at 23:43

4770K @ 4.7 - 780 @1137/1502

Without

With

https://imageshack.com/i/f191ftp

Sugarhell · 15 Mar 2014 at 23:48

Orangey said:
That image has been debunked, if your starting mesh was designed with 60 triangles in mind then 60,000 means you start over and not try and gild a turd. I think Nvidia do have a leg to stand on as far as pushing boundaries but not with tesselated water hidden under the map out of sight.

I just add the graph to see the difference between 6k triangles and 60k.Nothing else. Even if you make a mesh with 60k triangles going to 120k will increase by 5-10% the IQ.No more.But going from 60K to 120K is a huge amount of power

Also if you have experience with cry engine when you start a level its only water terrain. Thats why it seems that the water is over tesselated when its not. Its just the base terrain of cry engine

shankly1985 · 16 Mar 2014 at 03:12

Silent_Scone · 16 Mar 2014 at 05:34

Defy Belief said:
I'll do this.

Wouldn't it make sense to do a bench without tessellation then one with? And work out the percentage it impacts the card having it turned on?

Pretty much or else you might as well not bother lol. Unigine will still favour NV cards. Valley for instance.

http://www.ozone3d.net/benchmarks/tessmark/

Unfortunately tessmark is OpenGL, something else AMD cards are unbalanced in due to NVs own extensions. Can't win lol.

Kaapstad · 16 Mar 2014 at 08:38

Stock 290X @1000/1250

4930k @4.7

With extreme tessellation

With tessellation disabled

Increase in FPS 35.45%

I will be back in a bit after I have tried the same on a single Titan.