NVIDIA GF100 & Friends speculation

[Andrew Lauritzen]

It's quite a technical challenge, though. Until a triangle's vertices have passed through DS, you can't know precisely if the triangle's possibly occluded.

It's becoming clear to me that just raw tessellation/displacement in the graphics pipeline with no high-level culling based on bounding box guarantees or otherwise is going to be a poor use of hardware resources. Unsurprising really, but it casts doubt onto the utility of a naive benchmark implementation.

Though if high density is required to produce nice specularity, for example (which would be anti-aliased by MSAA) then high density is required even in areas of low/smooth displacement.

You never really want to tessellate smooth surface interiors... these cases can all be handled more efficiently with shading (even with higher-order normal interpolation if required).

Yes yes. But we know nothing about GF100's DC performance. All this DX11 downplaying is a result of GF100 tessellation performance.

No one is downplaying DX11... everyone loves it. What people are questioning is the realism of the tessellation workload in the one benchmark.

AMD has the only tessellation hardware for developer in the last 6 months. It seems normal that all games with tessellation show not very much tessellation (AVP or Dirt 2).

AvP has plenty of tessellation... the aliens for instance are fully smooth and tessellating them more would just be a waste (perhaps Uniengine-style).

Well simple subdivision isn't what makes tessellation useful. It's the displacement mapping that happens on top of that which really makes a difference.

Not true - tessellation is arguably even more important as a means of continuous LOD (*removing* polygons efficiently) than adding detail.

 

[Razor1]

Not true - tessellation is arguably even more important as a means of continuous LOD (*removing* polygons efficiently) than adding detail.

Well it goes both ways, this is from a design presepective you still have similiar triangles/mesh then before since the detail can't be cut down more then a certain amount.

what you said is true as well from a graphical point of view what Trini stated is true too.

 

[TKK]

This is the 470 cooler we're talking about right, with rumoured maximum board power of 225w - lower than the GTX 280? I guess I'm missing why a powerful fan and more elaborate heatsink than what we saw on the GTX 280 is cause for alarm. Or are you guys trying to say that a 225w 470 board will dump more heat than the 236w TDP 280? If I'm not mistaken that sort of defies the laws of physics.

Well, according to this http://gpucafe.com/2009/10/modern-gpu-failure-rates-1h09/, the failure rate of the GTX 280 clearly exceeded all other cards of that generation, and iirc the GTX 280 is much hotter than the GTX 260 under full load. Well that, and TSMCs 40nm is still being rumored to be more leaky than previous processes.
I think it makes sense for Nvidia to use a better cooling solution than on the GTX 280, even if the 470 consumes <225W.

 

[trinibwoy]

Not true - tessellation is arguably even more important as a means of continuous LOD (*removing* polygons efficiently) than adding detail.

Tell that to all the people waiting to see what all this tessellation hype is about. People don't want more efficient polygon handling. They want prettier models. Tessellation is of no use to us if it does not raise the bar on geometric detail.

 

[mao5]

Look at the RAM. It is a 470.

actually it's 480, you can type 470 in your photoshop, then blur it, to see if it get matched with the 470's blur image in the warhead bench picture.

 

[Razor1]

Well, according to this http://gpucafe.com/2009/10/modern-gpu-failure-rates-1h09/, the failure rate of the GTX 280 clearly exceeded all other cards of that generation, and iirc the GTX 280 is much hotter than the GTX 260 under full load. Well that, and TSMCs 40nm is still being rumored to be more leaky than previous processes.
I think it makes sense for Nvidia to use a better cooling solution than on the GTX 280, even if the 470 consumes <225W.

leakage doesn't constitute increased faluire rate of a chip, if I'm not mistaken.

 

[trinibwoy]

actually it's 480, you can type 470 in your photoshop, then blur it, to see if it get matched with the 470's blur image in the warhead bench picture.

Yes, that's a way more reliable method than simply realizing the unblurred memory capacity ends in 0 and not 6.

 

[mao5]

Yes, that's a way more reliable method than simply realizing the unblurred memory capacity ends in 0 and not 6.

also don't forget photoshop

 

[Alexko]

Perhaps, but that doesn't answer the question if why it should be harder to keep it cool in the first place (which is what I believe neliz and GZ007 were getting at). Or in other words given a particular level of power consumption how do you define an inherently "hotter" chip?

I think the point Dave was trying to make is that if GF100 needed to be kept at 70°C to remain within its power envelope while GT200 only needed to be kept around 90°C, then it would need a much more powerful cooling system.

Edit: whoops, mczak was way faster.

 

[Andrew Lauritzen]

Tell that to all the people waiting to see what all this tessellation hype is about. People don't want more efficient polygon handling. They want prettier models. Tessellation is of no use to us if it does not raise the bar on geometric detail.

It obviously does that too, but the key is that it *normalizes* it across the screen, and in-so-doing allows more efficient GPU execution. And remember, performance and quality are directly interrelated... performance gains can always be traded back for more quality.

Reduction/elimination of LOD popping is of direct visual impact to all users though. Most reviews I see mention these issues (or lack of them) in games directly!

 

[jimbo75]

Interesting results. It'll be interesting to see the results for dragon close-up with/without tessellation. Have to admit I'm surprised by the 8xMSAA performance - with so few triangles in the benchmark, overall and assuming the engine sensibly only uses MSAA where/when it's needed.

It's a shame the B3D article doesn't show HD5770 results as well. Oh well.

Jawed

Dx11, High Shaders, Tesselation Off, Anisotropy 1x, AA Off, Full Screen 1680x1050

Total fps = 39.5, total score = 995
Average at dragon = 43
Low at dragon = 40

Dx11, High Shaders, Tesselation On, Anisotropy 1x, AA Off, Full Screen 1680x1050

Total fps = 24.9, total score = 628
Average at dragon = 15
Low at dragon = 13

Dx11, High Shaders, Tesselation On, Anisotropy 1x, AA 8x, Full Screen 1680x1050

Total fps = 15.6, total score = 394
Average at dragon = 12
Low at dragon = 8

Dx11, High Shaders, Tesselation On, Anisotropy 1x, AA Off, Full Screen 1680x1050, GPU overclock to 950mhz, memory overclock to 1300mhz

Total fps = 27.5, total score = 692
Average at dragon = 17
Low at dragon = 14


----

I noticed on all counts that the low fps point at the dragon happened once or twice maximum, it didnt seem to stay at the low point for more than a frame just looking at fraps. It shouldnt matter but the cpu is a 3.6 ghz Phenom II.

 

[ShaidarHaran]

That 12V/1.8A fan will eat 20+W alone. Thats a very powerfull fan for 70mm.

This presumes the fan would be running at 100% load, which is a rather unlikely scenario. Of the recent NV hardware I've owned, I've rarely seen the fan speed exceed 40% while gaming, but approaching 50% when running FAH. This ranges from a 9600 GT at the low end to a GTX 285 at the high end (with a couple G92 and other GT200b cards in between).

 

[aaronspink]

I wish changes as big as gt200->fermi could be carried out in a matter of few months.

The sad part? Not that big of a change, double the SIMDs, add another piece to the front end, spin spin spin.

 

[GZ007]

This presumes the fan would be running at 100% load, which is a rather unlikely scenario. Of the recent NV hardware I've owned, I've rarely seen the fan speed exceed 40% while gaming, but approaching 50% when running FAH. This ranges from a 9600 GT at the low end to a GTX 285 at the high end (with a couple G92 and other GT200b cards in between).

The cards need to operate within a given range of temperatures. They dont want to exchange the cards just because someone is running the cards in a burning hot summer day in a room without air conditioner. Its a reference cooler design and they had numbers and reasons to use this cooling over a much weaker fan.

 

[trinibwoy]

It obviously does that too, but the key is that it *normalizes* it across the screen, and in-so-doing allows more efficient GPU execution. And remember, performance and quality are directly interrelated... performance gains can always be traded back for more quality.

Yep, all I'm saying is that simply applying a LOD system to today's models isn't going to benefit anyone much. We need to see a tangible increase in the maximum LOD as well and displacement mapping is going to have a much bigger impact than simply smoothing out pointy heads and elbows.

Reduction/elimination of LOD popping is of direct visual impact to all users though. Most reviews I see mention these issues (or lack of them) in games directly!

Well LOD popping today isn't even a matter of triangle density. We have entire boulders snapping into view in front our eyes so the problem is far worse than just the lack of tessellation

 

[Jawed]

The sad part? Not that big of a change, double the SIMDs, add another piece to the front end, spin spin spin.

On that basis Larrabee should have been here back in, ooh, 2005? Ye olde Pentium something or other plus a common or garden 16-way ALU, some cache and a bit of DXT decompression hardware. Not exactly rocket science, is it?

Jawed

 

[mao5]

480:

http://bbs.pczilla.net/attachments/month_1003/10030905531ec384e32cae19d5.jpg

5870 1GHz/1250MHz (core/mem)

http://bbs.pczilla.net/attachments/month_1003/1003090553cfa7405ba2e6364c.jpg

same setting as 480 in that video with Tessellation on.

haha, again, sorry Razor1

btw, anyone has the screenshot or video of 480 with Tessellation off at this place?

 

[Razor1]

480:

http://bbs.pczilla.net/attachments/month_1003/10030905531ec384e32cae19d5.jpg

5870 1GHz/1250MHz (core/mem)

http://bbs.pczilla.net/attachments/month_1003/1003090553cfa7405ba2e6364c.jpg

same setting as 480 in that video with Tessellation on.

haha, again, sorry Razor1

you do realize there is very little tessellation going on in that scene right outside of the building on the right side, there is nothing else there , the grass doesn't tessellate and all the other objects are too far away for any serious tessellation on them. Oh you edited for a picture of the area, use common scene they are using an adaptive tessellation shader based on depth, and FOV, what parts of that scene needs tessellation. Its obvious the grass doesn't, they are just planes, you can see the pixels on the edges of the grass, if they were polygons you will get aliasing, then look at the light tower on the right side, you have an edge of the light tower, that has aliasing but all in a single smooth line, no tessellation needed there. Trees are made out of planes, for the most part, so no tessellation there, so on and so forth.

 

[Jawed]

Nice numbers - it's amazing how simple stuff like this testing has fallen out of favour.

Dx11, High Shaders, Tesselation Off, Anisotropy 1x, AA Off, Full Screen 1680x1050

Total fps = 39.5, total score = 995
Average at dragon = 43
Low at dragon = 40

Dx11, High Shaders, Tesselation On, Anisotropy 1x, AA Off, Full Screen 1680x1050

Total fps = 24.9, total score = 628
Average at dragon = 15
Low at dragon = 13

Those minima are so close to HD5870 I suppose it indicates a significant bottleneck in fixed function hardware, i.e. some combination of TS and setup - though setup should be fairly minor if it's only the bottleneck 15% of the time in HD5870.

It still bugs me, though, how so few triangles with tessellation on can constitute a significant bottleneck.

Dx11, High Shaders, Tesselation On, Anisotropy 1x, AA 8x, Full Screen 1680x1050

Total fps = 15.6, total score = 394
Average at dragon = 12
Low at dragon = 8

The average at the dragon is 80% of AA off framerate, yet overall performance is 63%. And the low, 8, is 62%. I suppose, again, the average at the dragon seems to imply that fixed function geometry hardware is dominating.

I wonder if it could be something as simple as post transform vertex cache? Fermi has heaps of L1/L2 to buffer this data - but I presume an old-fashioned GPU with a piddly lickle PTVC is continually evicting vertices way too early, requiring them to be re-processed by VS/HS/TS/DS.

Dx11, High Shaders, Tesselation On, Anisotropy 1x, AA Off, Full Screen 1680x1050, GPU overclock to 950mhz, memory overclock to 1300mhz

Total fps = 27.5, total score = 692
Average at dragon = 17
Low at dragon = 14

The average is tending to scale with GPU clock and I suppose the low is bound by bandwidth (which has increased by less than GPU clock).

----

I noticed on all counts that the low fps point at the dragon happened once or twice maximum, it didnt seem to stay at the low point for more than a frame just looking at fraps. It shouldnt matter but the cpu is a 3.6 ghz Phenom II.

You should be able to change from fly through mode into free camera mode, like in Petersen's video, and see if you can really home in on the absolute worst case frame

Jawed

 

[mao5]

you do realize there is very little tessellation going on in that scene right outside of the building on the right side, there is nothing else there , the grass doesn't tessellate and all the other objects are too far away for any serious tessellation on them.

do you notice my "btw"? and you admit there's indeed a tessellation in the scene, so I think it represents a common DX11 game scene with tessellation.