Triangle rate vs Vertices rate, equals or not?

[vlask]

On reworked site i'm trying also show Triangle Rate count in milions triangles. Problem is that new cards (from DX7) are reffered only with Vertices rate instead Triangles.
Maybe i'm wrong, but vertices rate should be same as Triangle rate, just done by T&L unit.

Question is, can i write vertices rate into Triangle rate field?

Interesting thing is, that nvidia at Quadro FX cards show Triangle rate, but at gaming cards Vertices rate and gaming cards have way more higher numbers (around 80 vs 220+).

 

[OpenGL guy]

On reworked site i'm trying also show Triangle Rate count in milions triangles. Problem is that new cards (from DX7) are reffered only with Vertices rate instead Triangles.
Maybe i'm wrong, but vertices rate should be same as Triangle rate, just done by T&L unit.

Question is, can i write vertices rate into Triangle rate field?

Interesting thing is, that nvidia at Quadro FX cards show Triangle rate, but at gaming cards Vertices rate and gaming cards have way more higher numbers (around 80 vs 220+).

Triangle rate is about equal to vertex rate when you are working with primitives like triangle strips and fans. In those cases, you just need 1 new vertex to make a triangle from two old vertices. With triangle list (each triangle is composed of new vertices) performance may go down. Using indexed primitives can help here as long as you reuse indices that are still within the vertex cache.

This information may be stale, I haven't worked on graphics in ages.

 

[3dcgi]

Your information isn't stale. Triangle and vertex rate are used interchangeably. At least for AMD hardware and likely for others too.

 

[mczak]

It is not really technically accurate however.
In case of nvidia chips for instance, gm107 (kepler/fermi work much the same just different numbers) each smm can output 1 vertex or prim per clock apparently (at least I think at this level it's really vertex limited too, so if you'd use tri list it would be 1/3 vertex hence 1/9 prim per smm). So with 5 SMMs that's 5/3 prims per clock - but the chip only has one rasterizer so effectively it's just one primitive per clock. However, the rasterizer should be unaffected by the type of primitive you use, so if you'd use tri list, it would be 5/9 prims per clock as limited by the SMM, not 1/3. And if you count culled prims too, then it isn't limited by the rasterizer, and the chip can indeed sustain the full 5/3 rate (for tri strips).

 

[Exophase]

Either term is ambiguous as to which bottleneck it's referring to. Could be triangle setup, vertex shading, culling.. who knows.

 

[vlask]

Ok, since both are theoretical values and somehow comparable, i will use only one field for them. Chip makers showing one or another number, never seen vertices/triangle rate both at one chip, so i assume that they mean same thing, just with different naming.
Want to use these numbers for generating graphs (not tested yet, not sure if it works) - would be nice to have graphs (auto generated) with numbers (fill rate, texel rate, triangle rate, memory bandwith) about almost all 3D cards since 1996.

 

[Putas]

would be nice to have graphs (auto generated) with numbers (fill rate, texel rate, triangle rate, memory bandwith) about almost all 3D cards since 1996.

Like this? Quite intensive, and hard to see differences among oldest cards.

 

[Blazkowicz]

This thing is hilarious, I hope the page's author is doing something wrong with javascript and such. (I mean shouldn't there be small database crunching server-side, and then client-side drawing text and rectangles for less than one second of CPU time?)

Offering the choice of a log scale and/or a date range would fix the latter problem.

 

[vlask]

Like this? Quite intensive, and hard to see differences among oldest cards.

Yep like that, but with numbers. Those you linked are uselles. Even % of slowest card are better than this.

 

[Putas]

Okay, the demanding thing was animation of hundreds of bars. I disabled it, added limiting year and displayed values in the chart. There seems to be 8k pixel limit for flash window, hehe.

 

[Lightman]

Okay, the demanding thing was animation of hundreds of bars. I disabled it, added limiting year and displayed values in the chart. There seems to be 8k pixel limit for flash window, hehe.

Much better now

My i7 4770k can now cope reasonably well wit these enormous tables, but the info in them is priceless!

Thanks for doing it!

 

[Laa-Yosh]

Practical results also depend on the art content. Additional bone weights, UV coordinates, or breaks in UV space, or in shading etc. etc. can all affect the actual triangle/vertex ratio of the characters, props and level geometry. There's no general ratio and that is why HW manufacturers prefer to use vertices per second.