How powerful is Model 3?

[Gjay]

Ahoy!

We all know that Model 3 is a beast when it comes to memory and bandwith with all the masked ROM and RAM.
But how many polygons per second can it actually do at it's ~60fps?

In the product overview, it says that the Real3d-pro 1000 is capable of processing 16.500 polygons for each frame at 60fps, that would be 990.000 polygons per second.
(Those polygons already are textured, shaded, anti-aliased,fogged, illuminated, translucent and Z-buffered.)

That would put it higher than System 22, PS1, Saturn or N64, but much lower than Dreamcast (where DoA2 as one of the games with most complex geometry does 2.1 million).


Is this 990.000 numer true for the implementation in Sega's Model 3 as well, as Sega made some changes?

What about the later Step 1.5 and 2.0/2.1 revisions?

[Bart]

This is really hard to say! It wouldn't be that hard to instrument the code to print out the number of polygons processed each frame to at least determine what Model 3 games were throwing around. I don't think there were any substantial modifications to the Pro-1000 ASICs used on Model 3 boards.

Theoretical limits are always tricky because we don't know how they were computed. For example, let's say the geometry processor can transform and light 16,500 polys in 1/60th of a second. Can that many polygons then be rasterized and depth tested? Can that many be read from VROM per frame (one nice advantage Model 3 has is that the scene graph can be uploaded and forgotten by the CPU, so CPU->GPU bandwidth is irrelevant).

Can you remind me on Wednesday next week to print out the number of polygons per frame? Prepping for something that day and can't do it now.

[Gjay]

Please do that! I would be interested in the results (even if I can't test it on my own as I can't afford a graphics card).
I put it in my calender for wednesday.

The documentation says those 16.500 polys are rasterized and z-buffered. But those numbers seem so low to me.

As a big fan of the Dreamcast, for example, I can say that this system is more limited by available memory (next to fillrate and bandwidth) than it is by vertex calculations. (Maybe that was the time when polygons per second lost their meaning in determining how capable a system was, as it was not the limiting factor anymore.)
Nevertheless a few games did more than 2 million pps per second, and theoretically even more was possible.

Now, games like L.A. machine guns or Daytona 2 still look geometrically more complex than any Dreamcast game i know of, and while looks maybe deceiving, they certainly don't look like only pushing half as many polys around...so I'd like to know what's going on.

[gm_matthew]

I decided to cook up something myself that tracks the total number of polygons (and also non-repeating vertices) rendered per frame, and the highest amount that I've seen is 24,582 polygons and 66,610 vertices in a single frame during the attract sequence of Emergency Call Ambulance.

However, the real Model 3 is not actually rendering all of these. If you watch this video closely at around 3:06 you can see a barrier in the distance that is rendered at first but eventually disappears because the hardware can't render it before the start of the next frame; the Model 3 stops rendering if it runs out of time in order to maintain a constant 57.5 fps.

I bring up the number of vertices because it is arguably more relevant than the number of polygons. A 500 polygon model with no reused vertices will take much longer to perform T&L calculation for than a 500 polygon model with efficiently shared vertices. Also the Model 3 can render quads which have four vertices compared to three for triangles, plus quad strips require two new vertices per quad while triangle strips only require one new vertex per triangle.

From my personal testing, I estimate that the Model 3 is actually able to calculate around 45-46k vertices per frame, which in ECA at least is enough for around 16-17k polygons per frame or roughly 1 million polygons per second. If each quad is split into two triangles, the figure becomes ~1.8 million triangles per second which is not too far off what most Dreamcast games can achieve.

Worth pointing out that the Model 3 step 2.x video board runs at 66.666MHz while the Real3D Pro-1000 runs at 50MHz. I also found this document that claims the following:

POLYGONS PER CHANNEL:

Throughput averages 750,000 polygons per second per R3D/PRO-1000
unit. Polygons are fully rendered, textured, shaded, fogged, anti-
aliased, translucent, three- or four-vertex objects in the scene.

Considering that the step 2.x video board is clocked 33% faster than the Pro-1000, one million polygons per second sounds about right.

[Gjay]

Could you check Daytona 2 and L.A. Machineguns?


I agree vertices are more important.
I only asked for polygons/sec because for comparability reasons, since there's lots of hardware from those days where there are only polygon rates made known.

For triangle polygons, vertex count and polygon count should be almost the same, as you would use triangle strips almost all the time (even less if you used indexed geometry, but at least the Dreamcast did not support that).

For quads, two vertices should roughly equal one quad, which in term should be roughly equal to two triangle polygons.


Now, I wonder why we got two different figures here... 750.000 pps in your file and 990.000 in mine, both for the Real3D Pro-1000...


See page 17:
https://retrocdn.net/images/7/75/Real3D ... iption.pdf

[gm_matthew]

Daytona 2 is pretty efficient with polygons, peaking at around 14k polygons and 42-43k vertices during a race around the Expert track. L.A. Machineguns however managed to beat ECA with a peak of 22,638 polygons and 68,863 vertices during the attract mode, stage 1 demonstration. I ought to check videos of it running on real hardware to see if I can spot any models disappearing.

I didn't know about indexed geometry, guess you learn something new every day

EDIT: Unfortunately the attract mode includes randomness so I can't do direct comparisons

[Gjay]

I always thought that the beginner's Track of Daytona 2 - Battle on the Edge had some pretty elaborate geometry, plus the 40 cars.

The first track here, I mean:
https://www.youtube.com/watch?v=Gp-hMCUT9F0

(That track was removed in the power edition and replaced by a much more boring track, IMO)

BTW, ECA looks damn impressive at night courses!



I wonder why the vertex count is so much higher than the polygon count in your observations.
There should be roughly about twice as many vertices than polygons - but in these examples, it is way more than that.

Are they using so many individual quads? For what reason?

[gm_matthew]

The beginner track of BOTE is actually the least intensive in terms of polygon count, peaking at 12.7k polygons of which 10.5k are quads and 2.2k are triangles.

I know for a fact that most of the player car model is stored in polygon RAM rather than VROM and all of the polygons are individual with no shared vertices; I'm guessing the developers found it easier to add damage to the car this way. But even in general, Model 3 games aren't really optimized that well in terms of models reusing vertices.

[Gjay]

Polygon strips usually already happen when you create the 3d-models.
There should be no need to optimize this later, they share vertices automatically at the modelling process.

I mean they have a TON of memory (RAM, VRAM, CROM, VROM and more) on the Model 3, so you're right that they can get away with not optimizing as much easily;

but usually there is no need to do this for polygon strips, as this already is a normal process in model creating and models get exported this way, too.


Keeping the cars in RAM for damage purposes makes sense, tho.