Should the SEGA DC and Saturn have launched with these alternative designs?

N64 numbers are with full effects. Rendering quality like we have today, essentially, unlike what PS1/Saturn produce. N64 also has dedicated geometry hardware in the RSP unlike Saturn. I've seen ERP say that N64 really was capable of more geometry complexity than PS1 regardless of the stated PR numbers. They just needed to be able to make decent microcode for the RSP.

Dreamcast has no dedicated geometry hardware (other than usual triangle setup). The SH4 is designed to do it well, but the SH4 is also going to be spending a significant part of its time working on other things too. DC is like a PC in this way. So those high numbers are probably way outside the realm of practical game performance.

Real3D Pro 1000 chips have T&L processors. So while the arcade machines use a gimpy PowerPC 603e, the CPU isn't doing geometry processing. That is undoubtedly why the Model 3 games have much more complex geometry than the 3DFX arcade machines that had their MIPS R5xxx CPUs doing geometry. There's no doubt that Model 3 was a geometry monster for a couple of years.

 

Wasnt the PS1 capable of 300,000 polys with full effects on?

 

Only if there's a game moving 300k polys/s. I don't really know what the most complex PS1 games are. But obviously PS1's "full effects" has its own meaning. It doesn't exactly have many effects/features.

N64 even has full time edge AA in many games. Very clean, but too bad about the low res textures. PS1 textures would probably look very similar if you threw on bilinear filtering.

 

Nope.

PS1 has the following specs as far as I remember:

GTE: 1,500,000 million lines/sec
GTE: 500,000 raw polys/sec transform rate
graphics chip: 360,000 flat shaded polys/sec displayed/rendered to screen
graphics chip: 180,000 textured, gouraud shaded, lit polys/sec displayed/rendered to screen

That 180,000 figure is the important one. It's not even close to 300,000

Namco's high-end (for the time) System 22 board powered by DSPs and Evans & Sutherland TR3 is capable of 240,000 textured, gouraud shaded, z-buffered, lit polys/sec with all other effects on. If PS1 could do 300,000/sec it would be, in some ways, more powerful than System 22 which is absolutely not the case.

No not quite. Dreamcast outguns MODEL 3 in MOST aspects. Not every. Some people within SEGA's own AM divisions said that MODEL 3 was still somewhat superior to Dreamcast/NAOMI in some areas. I think anti-aliasing was one thing. AA comes for free within MODEL 3's stated spec. On Dreamcast it costs more.

 

I've read many times that N64 can do 600,000 polys/sec raw. That's probably flat shaded, if not just the transform rate.

True.

Exactly.

The high-end Real3D-Pro/1000 chips are true GPUs. They have their own on-chip geometry engines. We did not see even close to MODEL 3-ish performance in 3DFX-based arcade games until the Voodoo3, (i.e. Midway's 'Graphite' board powered by a PentiumIII @ 733 MHz and an Quantum3D/ 3DFX Voodoo 3). By then, CPUs supplying the T&L were fast enough to reach near MODEL 3 geometry performance. Although Voodoo3 still didn't have the rendering quality of MODEL 3. Things like AA were always better on MODEL 3. Even Dreamcast didn't reach it, even though Dreamcast did surpass MODEL 3 in many ways.

 

aa almost free on model 3? i doubt it. On powervr2dc on the other hand, it was cheap. I have first hand experience of this from my own coding efforts. With 3 million triangles per second covering 80% of the screen 2x true supersampling aa costs something like 20% more than non-aa rendering. And it needs no additional framebuffer space as well, let's not forget that, huge bonus.

Model 3 was just an immediate renderer, no way it could match powervr2dc's deferred approach, especially when you consider it was clocked significantly lower...66 mhz if i remember correctly. Not to mention the features it lacked when compared to naomi/dc. As for the geometry engine, the sh4 was faster than model 3's implementation and being a normal fpu as well quite more flexible and with a direct path to powervr2dc so that you could send your shaded results direclty for rendering (i assume model 3's geometry engine was something like opengl's fixed vector transform engine). And it was monstrously more expensive too. There was a real reason why sega abandoned the Model X architecture and went with sh4/powervr...performance and price.

a list of key features that you got in '98 for powervr2dc's price:
- deferred rendering
- no wasted fillrate on hidden transparent triangles
- per pixel transparency sorting (just think what you have to do nowadays to get this working properly)
- cool accumulation buffers with which you could do multitexturing+ effects not possible with other architectures of the time, actually you could cut any shape you liked from the already rendered contents of the hidden accumulation buffer and apply them to the front buffer, very cool.
- 1024x1024 textures
- punchthrough textures for super fillrate efficiency where applicable
- trilinear filtering
- anisotropic filtering
- vq texture compression
- high quality 16 bit yuv textures
- normal mapping (nobody used it, i am actually convinced devs did not know how to use the dc because 3d production pipelines were still new and they were happy using just one texture+one color per vertex lol )
- cheap 2x supersampling aa with no extra framebuffer storage cost
- 32 bit depth testing
- internal 32 bit rendering at all times, output to high quality dithered 16 bit if you wished to keep framebuffer size down
- volume modifiers
- plenty of blend modes
- comprehensive list of flicker free interlaced rendering modes
- vga output
- pretty sure i am forgetting some + the same chip contained the system's memory interface and other stuff...

How can model 3 even compete with something like the above that cost just a fraction of the price. It is absurd, when for the price of a model 3 board sega could put together a multi-naomi solution that could outperform anything up until 2000 with _ease_. On a tripping note, if dc was allowed to cost as much as the ps2 i think even that purely brute force architecture would have a really hard time even managing to keep up.

Also, 3+ mpps is doable on naomi for game scenarios, the only reason dc cannot do it is because there won't be any space left for textures. I have a 3 mpps demo that i wrote myself but consumes the cpu totally. However it is NOT the optimal approach, i will try that one when i have some really relaxed free time (switching from raw vertex arrays to indexed ones with smart cache management and plenty of hand written pipelined sh4 asm) and i am quite confident that performance results will go through the roof . Anyway, keep in mind that the most intensive games like DOA 2 did not go over 0.5 mpps....so dc's potential was left untapped.

This thread is so tripped out, model 3 is not better that naomi/dc, end of story

 

Oh ok.
Do you have any idea which games pushed the most polygons with full effects on each platform (Saturn, PSX, N64)? It would be interesting to see how well they performed in practice and how well they were represented on screen by actual games.

 

I'll take SEGA AM division's word that MODEL 3 was superior to Dreamcast/NAOMI in some areas. Even though Dreamcast/NAOMI was, overall, much superior to MODEL 3 in most aspects.

Look the CPU clockspeed on MODEL 3's 603e has almost zero impact on how good MODEL 3 was. It's like comparing it to the Apple Pippin which used the same CPU. It's all about the Lockheed Martin Real3D/Pro-1000 GPUs vs the SH-4 + PowerVR2DC / CLX2 graphics chip. End of story

 

The Dreamcast was capable of higher precision texture mapping, and had a 50% increase in maxmum real-time polygon throughput compared to the Model 3.

However, the Model 3 board was capable of pushing every one of its stated graphical effects in game whilst pushing the full 1 million pps onscreen at 60fps (Step 1.5 and higher). On the other hand, according to Sega themselves, with the DC, the more effects being pushed at once resulted in an impact in overall performance. For example, devs had to choose between having lots of specular and lighting, or a higher poly count with better texturing.

These are the main differences between the DC and the Model 3.

The Model 3's onboard geometry processor meant that the board could always display its maximum number of pps without compromising on effects. By contrast many DC Model 3 ports lack certain effects. The accomplished conversion of VT:OT for example only features character, and not environment reflections.

Model 3 used quads, which amounted to roughly a max of 2.2 million triangles compared to DC's 3.5 60fps max. To my knowledge, the Model 3 has never pushed more than 1.1 million pps on screen in any of its games. Although, with every stepping (step 1 to step 2 etc) the overall pixel fill-rate was upped allowing more overlapped tansparent polgons to be displayed.

Anyway I think that sums it up.

This was more down to the performance hit, than anything else. Even Sega's highest end Naomi titles lacked lower end (comp to bump mapping) visual effects the Model 3 could do for that reason. It's why Sega moved towards the Hikaru board, later replacing it with Naomi 2, which could do all of Model 3's effects without impacting on performance.

 

See, MODEL 3 was superior to Dreamcast/NAOMI in some areas. I knew it. I wish SEGA had launched a Lockheed-Martin Real3D based console instead of Dreamcast, even though I love the Dreamcast. The SEGA AM divisions loved MODEL 3 for a reason.

The MODEL 3 would've had two onboard geometry processors because it had two Real3D/Pro-1000s under the hood. Each one had its own geometry processor. Sure, the SH-4 had more floating point power and thus more geometry performance than the combined two Real3D/Pro-1000s but that's another matter. As you stated, MODEL 3 could run every single one of its effects/rendering features while pushing the stated amount of polys/sec. PowerVR2DC / CLX2 in Dreamcast could not do that.

 

Indeed. I was aware of the dual Pro-1000 set-up, just lazily reffering to the geometry processors, along with the other two chips, as one combined unit. If that makes sense.

Apparently Saturn VF2 only pushes 120,000 pps at 60fps, with gouraud shading making up for the lack of overall geometry. And according to an interview with Sega's CS team director way back in an issue of Sega Saturn Magazine, the Saturn version of Sega Rally featured a third less polygons than found the Model 2 original. About 200,000 pps at 30fps backed up with gouraud shading. The game also sported up to three colours per texture, whilst the Model 2 board could only handle monochrome textures, if I remember correctly.

Mario 64 pushes around 100,000 textured, bilinear filtered pps at 30fps with edge-based anti-aliasing.

As for the PS1, mostly likely Gran Turismo for polys at 30fps, and Ridge Racer Hi-Spec demo for 60fps.

It would be interesting to see the maximum number of pps taking into account 30fps titles and 60fps ones.

 

I'd ask ERP about World Driver Championship. I don't know if he'll remember frame polygon counts though. Boss really pushed N64 to the edge with that game.

I also wonder about Conker's BFD. It is in another class of visuals compared to most N64 games.

The early N64 games are gimped by inefficient SGI microcode, and developers who haven't figured out how to get more out of N64's not-so-ideal architecture.

 

Ridge Racer Hi-Spec demo was the most amazing thing on PS1 that I ever saw. I almost forgot about that.

 

so what could sega have done if the dreamcast was $300 instead of $200 ? Second sh-4 and more ram ?

 

Probably released something like a NAOMI 2 with a T&L chip/geometry processor and two PowerVR2DC chips, plus more RAM.

 

I think what would have been much more useful is a lot more RAM. 32MB maybe. There just wasn't enough RAM in consoles until Xbox came around. Just look at what they had to do to UT99 for DC.

Maybe spend some money on a controller redesign too. The DC controller needs more buttons.

 

My ideal console instead of Dreamcast, with an IBM G3 CPU and Lockheed "Real3D/500" GPU would have 48 MB for the system and 96 MB dedicated to graphics. A huge amount of RAM that beats the Xbox by alot. Of course it would not come out until 2000 or 2001 and be priced at $299. Keep in mind, that's a dream console ^__^

 

Well just imagine what a 128MB Xbox could have been like. The dev kits had 128MB. If you look at the mobo you can see the empty RAM pads. But they were already losing tons of money on it.

I would've also liked to see it with the originally planned Athlon CPU. Then it would've been scarily similar to the popular Athlon + nForce setup of the time.

 

Better yet, how about an Xbox with 512 MB RAM, like the Xbox-based Chihiro arcade boards

My ideal Xbox would've had a 1.5 GHz Intel Pentium 4 CPU with a custom Vector Unit for extra floating point performance, coupled with a full GeForce 4 Ti 4600 GPU @ 300 MHz and 512 MB of DDR2, launching in late 2002.

In this senareo, the PS2 would've also have been more powerful, using 180nm or 150nm tech, with about 3x the performance in the Emotion Engine meaning 18.6 GFLOPs, launching in late 2001. The graphics would've been Evans & Sutherland RealIMAGE tech, scaled up massively ala Graphics Synthesizer, thus meaning lots of pixel pipelines but also an equal amount of texture units and 12 MB eDRAM. Main memory would be 128 MB Rambus RDRAM and seperate 32 MB external graphics memory tightly coupled to the E&S GPU. Why do I choose E&S RealIMAGE tech for PS2's GPU? Because it's a rival of Lockheed Martin's Real3D tech. Also because NAMCO used older E&S graphics in their high-end arcade boards. Also because it would be good if Sony had a good graphics partner. Sega had Lockheed, Nintendo had Silicon Graphics.

 

FWIW a quad is not equivalent to 2 triangles except in certain circumstances - in the general case, a quad is only equivalent to a single triangle. <shrug>