NVIDIA GF100 & Friends speculation

[babcat]

So has the one triangle per clock barrier been broken?

 

[rpg.314]

So has the one triangle per clock barrier been broken?

By a mile. And then some.

 

[Bob]

So in anands article he says "8x the geometry performance of GT200". Is this from 2x the clock rate and 4x the triangle units? 8x would be quite the jump if that's true

The 8x number is normalized per clock (ie: assuming gf100 runs at the same clocks as gt200).

 

[rpg.314]

This is great news. IF it leads to better perf, then all this delay would have been worth it.

The ALU:TEX ratio seems to be 8:1, rather high. The caches are really saving bandwidth here. By routing a lot on inter-shader traffic via caches, GF100 is also doing away with many unexpected sources of bottlenecks in FIFOs. GF100 will outrun 5870, even though it has >1 TFlop less peak throughput. A full teraflop.

Demands on R9xx just got much bigger.

Let's hope that the thing that launches in March (A3?) hits it's clocks and is not a salvage part. Otherwise, having to wait till October for a full GF100 will probably mean that this thing ends up competing with R9xx.

 

[babcat]

By a mile. And then some.

So how many triangles can this chip produce per clock?

 

[AnarchX]

Anandtech states 64 texture-adresses and 256 texture-filters for GF100: http://www.anandtech.com/video/showdoc.aspx?i=3721&p=2
Is this number valid?

 

[rpg.314]

Anandtech states 64 texture-adresses and 256 texture-filters for GF100: http://www.anandtech.com/video/showdoc.aspx?i=3721&p=2
Is this number valid?

Doesn't make sense. It hasn't been done so far by any IHV and there is no mention of it in their graphics whitepaper.

 

[aaronspink]

Interesting:

So, 64 texture units but they are effectively clocked twice as high. Not too shabby.

At least thats how I read it, they are clocked at the shader hotclock?

according to the data available they are NOT clock at the shader clock or even twice as high. They are in fact clocked at 1/2 the shader clock which could potentially mean they are actually clocked SLOWER.

 

[rpg.314]

So how many triangles can this chip produce per clock?

In the best case, 16 I think. 1 for every polymorph engine.

 

[aaronspink]

Don't know how the clock domains are going to work out, only that the Texture Units do not operate at the GPU/ROP clock, they will be higher, if they match Shader clock I do not know yet. No information on clock speed domains.

According to Anand they are at 1/2 shader clock.

 

[Rys]

In the best case, 16 I think. 1 for every polymorph engine.

4, there's one for each block of four SMs. Texture happens at half the hot clock.

 

[babcat]

4, there's one for each block of four SMs. Texture happens at half the hot clock.

It seems NVIDIA has blown past the barrier!

4 polygons per clock. WOW!

 

[Ailuros]

http://anandtech.com/video/showdoc.aspx?i=3721&p=2
"Finally, texture units are now tied to the shader clock and not the core clock. They run at 1/2 the shader clock speed of GF100."

I wonder how much more texture performance its going to have than gt200. can't be much if it is indeed half the hot clock and only 64.

http://img197.imageshack.us/i/1001181407ac184361b265f.gif/
http://img197.imageshack.us/i/1001181407a1b9234747734.gif/

 

[Ninjaprime]

It seems NVIDIA has blown past the barrier!

4 polygons per clock. WOW!

Sort of. Appearantly for larger "normal" polygons its still 1/clock. Although I can't see using too many of those if you are getting near the setup limit anyway.

 

[DemoCoder]

One would also assume that there could be bottlenecks in pathological cases where say, all 4 setup units are producing overlapping triangles. As usual, there are always bottlenecks, the issue is the expected throughput for expected workloads.

One could say that for workloads with lots of large triangles, setup performance goes down, but then again, the total number of triangles is most likely smaller, likewise, for finely tessellated workloads, the setup rate is likely to hit peak 4tri/clock throughput.

 

[CarstenS]

One would also assume that there could be bottlenecks in pathological cases where say, all 4 setup units are producing overlapping triangles. As usual, there are always bottlenecks, the issue is the expected throughput for expected workloads.

So, what would Nvidia-PR call this? A Triangle-Virus? *SCNR*

 

[CNCAddict]

One thing that I'm not sure of. Are these "tessellated" triangles just as good as "normal" triangles. What downsides are there to tessellation? Also, if it's possible to create geometry from a normal map...then would that make it possible to add "real" bullet holes in games? If I shoot a door...could the 2 displacement maps on each side of the door converge to form a real, actual hole thru the door? I have been waiting for that ever since the cut scene in HalfLife-1 with the light bleeding thru the bullet holes in the air conditioning vent.

 

[Ninjaprime]

Essentially what they've done is divide up Ye Olde setup/tri unit into 4 mini setup/tri units, or so I interpret. File this under the "completely different" speculation.

I think what this means is that current games won't really care about it, as seen by their Far Cry 2 benches, but future tesselated games will see large benefits.

 

[CarstenS]

I think the optical effect would be quite easily achievable (in fact, it is already in some physx-levels for UT3 with whole walls). The larger problem is, getting the updated game-world info back to the cpu because if there's a real hole somewhere, you could shoot right through it.

 

[jimmyjames123]

I'm propably mistaken, but from what I have read, even prior to today, the Nvidia Surround would work with 19/24/19 setup and run them at native res where as Eyefinity you need 3 same size/res supporting monitors? Is that true, if you dont know can you find out? I think that would be a huge win for Nvidia if they could do that. Not everyone can afford huge displays be it cash flow or desk space, but 400 bucks for a 19/24/19 or even a 19/22/19 setup I believe more people could afford.

According to the GF100 whitepaper, gaming with multiple monitors in 2D (ie. non-stereoscopic 3D) will be supported for resolutions up to 2560 x 1600 across displays that share a common resolution. The whitepaper says that 3D Vision surround is supported across three of the same 3D Vision capable LCDs and projectors at resolutions up to 1920 x 1080.