RSX: 1.1 billion vertices/sec?

[BOOMEXPLODE]

430 MHz G70: 860 million vertices/sec.
550 MHz RSX: 1.1 billion vertices/sec. (extrapolated)

It looks like G70 can handle two vertices per clock cycle. That's funny because I remember the old banter on this forum about a 1 teraflop Cell and a billion polygons per second. It looks like atleast the last part of that is correct!

 

[Rockster]

You're confusing setup speed with vertex shader performance.

 

[nAo]

I think NVIDIA is basing their numbers on a 4 DOT4 instructions vertex shaders, the minimun one'd need to transform a vertex.
Since they have 8 vertex shaders they can trasform 8 vertices in 4cycles -> 2 vertices per cycles -> 430*2 -> 860 MVertices per second.
The same reasoning can be applied to RSX too..even if it doesn't make much sense from a real world application point of view.
CELL SPEs can also trasform a vertex in 4 cycles (even if they have to transform more vertices at the same time to hide fmadd latency) and since SPEs are clocked at a much higher clock than RSX I believe CELL could outperform RSX in this meaningless 'contest'.

 

[ralexand]

Anyone know how ATI came up with their 500 million number for the xeros?

 

[Shogmaster]

Anyone know how ATI came up with their 500 million number for the xeros?

It's the maximum triangle set up number. Xenos can calculate far more IIRC.

 

[ralexand]

Anyone know how ATI came up with their 500 million number for the xeros?

It's the maximum triangle set up number. Xenos can calculate far more IIRC.

Thanks, do you know how they come up with the max setup number?

 

[Shogmaster]

Anyone know how ATI came up with their 500 million number for the xeros?

It's the maximum triangle set up number. Xenos can calculate far more IIRC.

Thanks, do you know how they come up with the max setup number?

I think it's literally how many triangles Xenos can set up per second. Xenos can set up one triangle to draw per cycle.

 

[Titanio]

Xenos max triangle setup = 500m/sec.

RSX I guess would be 1.1bn if we can extrapolate from the G70 figure.

 

[nAo]

Don't mix setup engine throughputwith vertex shaders throughput, we don't have G70's setup engine performance figures

 

[Titanio]

Don't mix setup engine troughput with vertex shaders troughput, we don't have G70's setup engine performance figures

Ah, I see. Highly confusing! I guessed that was the setup figure.

 

[DeathKnight]

Is it vertices or triangles? If it's vertices then max triangle throughput is obviously much less. One vertice does not a triangle make.

 

[Titanio]

Is it vertices or triangles? If it's vertices then max triangle throughput is obviously much less. One vertice does not a triangle make.

With vertex sharing, it pretty much can. For figures like these, often the terms triangles and vertices are used interchangeably.

Think about it. You have 2 vertices. To make a triangle you add one vertex. To make another, you only have to add one more vertex. And so forth. Every new triangle may only cost you one more vertex, thus #vertices can = #triangles, pretty much.

 

[nAo]

Is it vertices or triangles? If it's vertices then max triangle throughput is obviously much less. One vertice does not a triangle make.

Umh...it depends
With post transformed vertices cache and indexed primitives one can have even <1 ratio between vertices and triangles!
I think most of the time with indexed primitives vertices/triangles ratio ranges from 1 to 1.5-2..

 

[Fafalada]

One vertice does not a triangle make.

Correct, one vertice does two triangles make. 8)

 

[Acert93]

As nAo was saying, Xenos triangle setup is 1 triangle per clock, 500MHz = 500M triangles/s.

The hardware tesselator unit can output 250M triangles/s.

I believe one of the devs had noted that Xenos, theoretically, if all the ALUs were dedicated to triangle transformation it was some insane number (like 6B triangles). Insane stupid number we will never see ingame (the dev was NOT suggesting we would either).

It has also been stated to expect a lot loss in general, butr on the reverse with non-trivial shaders the chip is closer to the theoretical than past chips. e.g. One dev had noted that Xenos could possibly be triangle setup limited in a game used Xbox1 type shaders.

Anyhow bits, poly/s, flops, etc... are only relevant in the big picture. The numbers themselves tell us very little without knowing the whole story about the graphics chip, the hardware setting it is being used, and the software expected to run on it.

 

[ralexand]

Anyone know how ATI came up with their 500 million number for the xeros?

It's the maximum triangle set up number. Xenos can calculate far more IIRC.

Thanks, do you know how they come up with the max setup number?

I think it's literally how many triangles Xenos can set up per second. Xenos can set up one triangle to draw per cycle.

Now I'm really confused. Does the triangle setup take place in the ALUs? Is triangle setup the same as vertex setup? Is the Tom Cruise Katie Holmes relationship a sham?

 

[Thowllly]

Is it vertices or triangles? If it's vertices then max triangle throughput is obviously much less. One vertice does not a triangle make.

No, but one hundred vertices can make one hundred triangles. In fact, one hundred vertices can make at least 198 triangles (maybe more, but I can't think of how). I belive that for most models, the triangle number is in fact higher than the vertex number...

 

[nAo]

Now I'm really confused. Does the triangle setup take place in the ALUs?

Nope, it takes place in the setup engine

Is triangle setup the same as vertex setup?

I think it is..but I dunno where you read about vertex setup..

 

[ralexand]

As nAo was saying, Xenos triangle setup is 1 triangle per clock, 500MHz = 500M triangles/s.

So since the G70/RSX can do 2 vertices per cycle then that's where we get the 1.1 bln number? Have we come full circle?

 

[DeathKnight]

Is it vertices or triangles? If it's vertices then max triangle throughput is obviously much less. One vertice does not a triangle make.

With vertex sharing, it pretty much can. For figures like these, often the terms triangles and vertices are used interchangeably.

Think about it. You have 2 vertices. To make a triangle you add one vertex. To make another, you only have to add one more vertex. And so forth. Every new triangle may only cost you one more vertex, thus #vertices can = #triangles, pretty much.

Yeah, I thought about that right after I posted :? Still, the Xbox's NV2A could really only achieve its theoretical maximum with triangle strips where pretty much every vertice was shared. I don't expect it to be any different with the RSX. Like others have said there are many limiting factors and the number doesn't mean a whole lot.