NVIDIA GF100 & Friends speculation

10fps or 3fps, it's a slideshow. This had better be a gtx 470, because if the 480 is losing to the 5870 in a TWIMTBP title nVidia might as well forget about releasing anything.

 

Look at the RAM. It is a 470.

 

IIRC, there were also gems like, "because of physics, our $129 gpu is faster than their $400 gpu. "

 

That si tied to shader clock in fermi.

I didn't know about that. Thanks for this.

 

IMHO, it is a dead heat. 5870's atrocious min fps is the only standout data point.

 

eagle eyed...

 

Good minimum rates for GTX470 in this bench. The cap on maximum FPS could indicate limitation somewhere in the raw fill-rate or frame-buffer bandwidth front.

 

I'm guessing due to 20% more video memory this is gonna be a repeating pattern. In those memory limited situations, 470 and 480 will pull ahead.

HD5870 with 2GB will prolly be a different story though. Overclocked to 1GHz and even GTX480 will have a hard time matching it.

 

Oh right. But again the extra memory on the 470 is a bit of an x-factor so it's too early to tell whether the higher min-fps is due to something else.

Yep, but I thought we were talking about DX11. That bit of comedy was about promoting PhysX IIRC.

 

So far, 470 (and to a limited extent 480) looks to be within reach of more RAM, more clocks part. Increasing bandwidth might be an issue for them however.

 

In gameplay crysis benchmarks, 4890 already was head-to-head with GTX 285 in higher resolutions:
http://www.pcgameshardware.de/aid,6...ectX-11-Preisbrecher/Grafikkarte/Test/?page=7
So, nothing new here.

 

sure, look here (again the same gif ):



Different res. but the overall effect should be the same.

 

Strange how the minimum of the 5870 is so low in ambush compared to the lows in your benchmarks for Hells Heart.

Are we looking at a single point in the ambush bench where the 5870 falters? Or is it a series of lows combined with a few high spots keeping the averages on par with the 470?

 

Without a frame diagramm it's nonsense to discuss about min frames.

 

Yeah, but check out the average and maximum framerates. That's not good enough for Fermi given it's size and price, as well as being half a year behind. One minimum frame stall from 5870 (probably due to ram) isn't really worthy of a red underline.

 

Here's how the 5870 behaves in 1920x1200, couldn't find that diagram in 2560x1600:



The increase in resolution might explain a very short drop to about 6FPS, but dropping to only 2FPS has to mean it ran out of RAM.

 

Aha. In theory tessellation lowers percentage of culling, because the number of triangles varies with distance to camera and "back faces" with lower triangle density are further from the camera. Also, tessellated stuff that's occluded by the dragon should have lower triangle density than the dragon, again due to LOD-adaptive tessellation.

13.6x primitives implies that there's only around 120,000 in the untessellated version, assuming the 1.6 million referenced in the B3D article. Fairly rough estimate though, because amplification varies with distance from camera. Despite that, 120,000 is paltry for a game scene these days, isn't it?

I agree.

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.

I think the question rests on how dense the triangle mesh needs to be in order to provide a reasonable degree of detail on the farthest-displaced "vertices", e.g. on the dragon's body or roof tiles. It seems like the mesh is too dense on the body but too coarse on the spikes, for instance.

The original triangle mesh (untessellated) is so coarse that tessellation isn't being adapted to the locality of required geometric detail. 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.

Jawed

 

Fake Photo--< The Device ID is from GTX 275 >

 

Plus wrong Texel-Fill. And not even close to the launch-driver.
Bad (paint) job, Marc!

 

DIRT2 bench leaked. The rest has already been exposed before. Look like they are probably from the the same source.