R300 - Final Specs?

[Gollum]

Not really. Most of the vertices are shared in a highly optimised mesh. I'd wager that the real time number would be somewhere between 200 and 300 million triangles per second.

Yup I oversimplified and forgot to consider shared points, my bad, sorry. I just checked the statistics on some of my models here at work and they average about half the tris (e.g. 2600 vertextes, 1300 polys), this numer can change depending on what and especially how you model but it should be a good average to assume 1/2 tris/vert...

edit: some models I checked were based on quads before I triangulated them, so they're more of a worse case and show a horrible tri/vert ratio, so I gotta appologize again. On some realtime models I built some time ago I could actually see a ratio of up to 3/2 tris/vert, so acually you can easily have even more tris than vertexes if the model is optimized properly.

Also, that 125mil figure from the GF4Ti is probably vertextes instead of tris too, so the R300 should have like 3x its triangle throughput if those 400mil are based on facts, which is quite an impressive leap, even if bandwidth can't keep up with computation power..

 

[Xmas]

You can easily inflate transformation numbers: multiply them by three. Each vertex goes through three spaces, so count the vertex transformation rate once for each space. Nobody said they would only count it once for all three spaces. Just a thought.

I guess you're talking about model/world/view space. If you look at it like this you could take any number, because no matter how many transformations/space conversions you do the final operation will always be just one matrix multiplication.

 

[Mummy]

Well, a Geforce 4 4600 can handle 30fps in Codecreatures benchmark, that benchmark uses 650MPoly in some scenes (and even if im not particularly excited about it, some effects are there and it looks good), and that's at 300Mhz with 2 VS.

I'd speculate we can approach 1.2MPoly per frame at 60fps with NV30 and R300 or 2.4M at 30fps

That's pretty amazing.

 

[hughJ]

"So the R300 does relatively the same number of vertices per second is what you are saying?"

no, I meant that you could compare the 2 numbers as apples and apples, since they are both talking about vertices..

 

[Bambers]

The gf4 4600 is 136Mvertices/s so if the r300 is 400M then that is about 3x faster.

 

[Colourless]

On a general humanoid of about 1000 verts, you will often be around 1 verts per triangle, if not better. Checking some of the quake 3 models, they are actually a bit better than 1 vert per triangle.

 

[CMKRNL]

I'd speculate we can approach 1.2MPoly per frame at 60fps with NV30 and R300 or 2.4M at 30fps


I'd say your speculation is a wee bit high. Using a triangle strip mesh in VRAM an NV25 will give you about 11-12M tri/s. At 30fps that equates to at best 400k poly's per frame. I suspect that the R300 and NV30 will get 20M+ tri/s giving us 600-700k tri's per frame @ 30fps. This is real world performance as opposed to some theoretical crap. This number may get higher if there is HW tesselation after the triangle fetch, but in real world scenarios I don't see the performance going much higher than this.

 

[jvd]

what do you make of this http://www.tommti-systems.de/go.html?http://www.tommti-systems.de/main-Dateien/chips.html

 

[Basic]

The unproffesional text hints that the image is crap. But who knows, maybe some OEM will have the bad taste.

 

[KimB]

A card running at 325 MHZ at the .15micron process and 107 million transistors has got to be "the hottest PC hardware available". I wonder if they have adopted some sort of way to keep it cool?

I would imagine that this card will have some honking fan a la geforce 4 ti cards do. If not ATI has certainly managed to do something that nvidia can't do. Create a 100 million + transistor that runs at .15 micron that still runs relatively cool. I suspect that they will have extra cooling attached given the incredable mass of transistors at .15.

325MHZ is a higher clock rate then the geforce 4 ti cards default, for some reason ATI manages to create cooler running hardware compared to what nvidia can muster. I would suspect that this is why nvidia is insisting on the .13 micron process for the NV30.

If ATI indeed has an external power supply, this is certainly going to be one incredibly hot video card.

 

[KimB]

Btw, here's an easy way to see how you can get roughly twice as many triangles as you have vertices.

Start by drawing an equilateral triangle (as best you can...doesn't need to be very good) on a piece of paper. It should be large, taking up most of the paper.

Then, draw a second triangle inside the first by connecting the midpoints of the three sides. You should end up with four triangles where previously there was just one. If you do this enough, you'll end up with nearly a 2:1 triangle/vertex ratio.

Here's a table of vertices vs. triangles with each step:

1: 3:1
2: 6:4
3: 15:16
4: 45:64
5: 153:256
6: 561:1024
7: 2145:4096
8: 8385:16384

As you can see, the series is quickly approaching a 1:2 ratio (one vertex per two triangles). Additionally, in a real 3D model, the series will converge even more quickly, since all triangles have shared edges (there are no border edges).

For those of you wondering how I came up with the numbers, I found a relatively simple series method of finding the number of triangles/vertices after an arbitrary number of steps. And if you examine the equations mathematically, it does become obvious that the series converges very quickly to a 1:2 ratio.

 

[ram]

Err I think the 125 million number is the theretical limit of the geforce 4 ti in vertices... According to this that number is attached to the geforce 4 ti 4400 the geforce 4 ti4600 does 136 million vertices.

The theoretical limit of a Ti4600 is 150 transformed MVert/s. The 136 MVert/s number is what can be achieved in practice, when everything is deactivated (rasterization ...). In the OpenGL.org forums was once a thread where they reached that number.