For me it was the PPC version of Tornado3D.  Being able to move the camera without falling back to wireframe?  Amazing.  Importing a basic low-poly model from another program and applying subdivision to create an awesome perfectly formed organic mesh?  About $900 if I recall correctly. :) 

Of course the 68k was still doing the real work, only using the PPC in a manner similar to the way GPUs are used now.   

Surely "Breathless", it was impressive on a 030/50Mhz with 320x256 res and it's Fog effects!

Really impressive was also the TBL demo "Tint". Not their best demo but the only one I got to run on my expanded A1200...

well i have here a compile of apoolgl
(http://www.planetjahn.de/apool/)
  for amiga 68k and mesa actually running on an a4k with software driver wazp3d by alain thellier. it doesnt work with hardware voodoo3 acceleration yet, since hyperions warp3d drivers for voodoo are broken. in software mode it is slow beyond usability, i expect it will not be playable with hw acceleration either but time will show.

also compiled some further opengl applications for amiga like billardgl, glexcess demo and few others, most of them are too slow or dont work with warp, but alain is working to improve it.

The worst thing with many 3D games and demos are their low framerate. That destroys the overall impression.

If it is possible to produce games for the Natami that run as smooth as for example Crazy Taxi on the Sega Dreamcast, we have something to look forward to. As I see it, it's better to make a game optimized for speed and smoothness instead of adding too much graphics to it - which slows it down. Yes, it's nice with good-looking games, but playability and game experience should always be first priority.

I always liked the vector demos with blitter effects! I think some of the fun disappeared when people started to buy accelerator cards for creating demos. I would say many of the new generation Amiga demos have a PC feeling added to them. Let's hope Natami demos will be different... :-)

The Star Wars pics looked to me like renders from Reflections or Imagine.  Nice highlights, but lack of bump mapping is quite apparent...  Not too bad though.

Genetic Species looked pretty nice.  I can not recall for the life of me the name of the German demo game I was really impressed by at the time.  I remember it took a long time to get running as the light maps had to be computed before playing.

I was always impressed by the software 3D tricks some demos used.  I always wondered why the PS2 couldn't do anything like that, and I guess it's because of the fixed function 3D pipeline.  I think the Dreamcast was one of the greatest systems ever.  Sword of Berzerk and Cannon Spike are two games that I found particularly good looking. 

Crazy Taxi was the bomb...  (You're goin' the wrong way!)   

One of my favorite demos just had a scene with some coins falling, but used some nice emboss or bump map trick in combination with a reflection map.  Those coins really looked to have a raised surface on them.  While no stranger to 3D, it was still the most believable real-world
surface I had seen outside of a raytrace/render.

I guess I'm wondering how flexible the 3D system will be.  I agree that games should be smooth, and for open world games with a lot of objects, this means minimal graphic effects.  Would still like to think that bump/normal mapping and post process filters like bloom could be used also when framerate is not crucial.

I should mention that I'm into real-time 3D, but more for film making than gaming.  I've never been into raytracing - my two favorite 3D packages for the Amiga (Caligari Broadcast and Tornado3D) both used scanline techniques. 

It seems Natami needs something like Arkham Development's Antiryad game development tool.  It might not be a huge stretch as an Amiga version was being developed as recently as a couple of years ago, demos are available here:

EXTERNAL LINK

Frontier Elite for the games, it is a big word with a lot of details, procedure generated, but still very big. The blitter use was pushed to the max, and the blitter was not powerfull enough for this game as soon as you put a big CPU accelerator.

For applications, it would be Caligari, on my A500, i've got a display for rendering like i could seen on the super expensive silicon graphics at the time (91-92) like the program wavefront.

Demo wise, i think the best is Lapsuus (Maturefuck) on 68060
EXTERNAL LINK  It was made by member of futuremark not long after the 3DMark2001 and you have some similar effects.
The limitation is the resolution, but it is still impressive adn it is a testimony of demo power :-D

A true triumph of design.  It wouldn't be any better if it had twice the polygons.  Maybe we could them back together to do a "remix" for Natami?

And I tought TBL were hard to beat... that is freaking awesome!
it's unbelievable it is running on a 60/70Mhz CPU, 3.5Mhz Chipset and 14Mhz internal BUS... 

Now figure out a NatAmi version of it!

The question was about the best realtime 3D, so technically, i'm impressed by the Lapsuus as it must use a lot of trick to use every single cycle available, but design wise, TBL's are far superior.

Do not forget that the image is computed in 160x100 resolution (16000 pixels), the blurring effect comes from the x2 scale to 320x200. The PC demos are 800x600 (480 000 = 30 times more pixels).
We need a 68060 at 50mhz x 30 = 1,5ghz to have the same demo in 800x600.
Quite bit a challenge to use the same effects in increased resolution on the Natami.

So this demo use every cpu cycle, but bus wise, it's not so impressive, it's still 320x200x8bppx60hz max for chipram bandwidth, the same as whatever VBL demo/game (60hz in NTSC) < 3.7 MiB/s
Max theoric chipram bandwith is still 56MiB/s

Each time i look at it, i'm always wondering if there is not an 24/16bit -> ham8 conversion on the last nature effects, even if they can achieved the same effects with precalculated palette, it would be simplier to add the conversion in the x2 scale function.

About the color palette of the nature effects (Literally JAWDROPPING), I think that rather HAM8 they used some copperlist trick to obtain more colors. Keep in mind that AGA Copper was able to switch palette registers each 2 - 4 lo res horizontal pixels depending on the used bandwidth (could be wrong but these are my memories). Seeing that the rendering resolution is 2X2 NTSC I guess they used that trick to pull out all the colors without the HAM8 limitations (can't see any colour crash)

But what do you mean with 14Mhz bus?
The chipmem ran at 3.5 MHz, right?

Are we talking about Starstruck?  I am halfway into it and already I am blown away.  I wish it were a game, as that type of high-tech sci-fi really appeals to me.  (There was a time when Japan made games and anime with a similar style.  Now they seem to have lost interest in that for a return to ninjas, fantasy and Western inspired mafia stories.)   

Still, they seem pretty close to me in terms of design.  If I weren't so into that exact type of stuff I don't know that I could pick one over the other. 

Is AGA+ true colour still planar?

Now to watch the rest...

In fact, the bus frequency of the AGA IS 14mhz and 32bits x2 (2 read by clock)
So in fact, in read burst mode max theoric bandwidth is 112MiB/s (if you consider you have a zero wait state, zero cycles latency addressing memory and a very looong read)

The chipset is only at 3.5mhz for "internal computation" (blittering, coperlisting,...) but the memory bus still is at 14mhz.

In fact the AGA chipset is just a resolution and color enhancement. The basic functions are still the same as the original Amiga 1000 (ocs chipmem bus is at 7mhz, and the chipset was already at 3.5mhz)

And as i read in this forum elsewhere, because the OCS could do 640x200 in 16 colors (4bpp), we could have got 320x200 in 256 colors (8bpp) among 4096 in 1985.
Shame that they do not have done an "MB mode" (Modular brightness, with the 3 last bits for luminance for example) if the 256 color register was too expensive relatively to the 32 color register.
It was not a bus limitation, and we could have got 4bpp dual playfield like AGA.

In general, you can recognize a TBL demo even for your first seeing, because of their particular "TBL touch style" (a bit like the melon design demos)
Starstruck focus on resolution and effects fading, but 320x200 is still blocky for today's standard.
Lapsuus has at least a "wow..." factor on colorfull effects for the nature part.

Do not forget either, that the 3D in the demos are mostly fake.
In the room with the rotating machinery and the hanging girl (around 2:10), i'm prety sure (99%) that all the coordinates of the machineries path are precalculated or transformated from one arm to another with symetry along the rotation axis.
You have still to draw the rendering polygons, and even, i'm not sure that there are no precalc on the gradient formulas.
They are some precal animation too behind the 3D or still panning background.

So you need a lot bigger CPU to render the same scene with no precalc, meaning that the demos achievement can not be achieved in real games.

One of the proof is that the polyhedra ball around 5:20 has only 30 polygons, the other parts of the demo seems to have a lot lot lot more, so why have only 30, if the computer is able to do a 200 poly ball, it's just a parameter for the coordinate generation.