Model 3 Step 1/1.5 and Step 2 Video Board Differences

[Gjay]

Hello again,

sry for asking a lot of technical stuff lately, but I don't know a better place to ask about this.
This time, I'd like to learn a bit more about the differences in the video-boards between Model 3 Step 1, Step 1.5 and Step 2.

1) To my understanding, the video boards are clocked differently (50 MHz for Step 1 and 66 MHz for Step 1.5/Step 2).
Now, as we don't have a single GPU here as we would have in later systems, what exactly is meant here? Bus frequencies, RAM, pixel processors, rasterizers, texturing units? Everything? : D


2nd question:
Board layouts for Step 1/1.5 and board layout for Step 2 look waaaay different.
There seem to be a LOT more differences then just clockrates.

If we compare:
Step 1.5:


and the board schematics for Step 2 (I took from the well-known thread in this forum):


Then it seems Step 2 has double the rasterizing and texturing units of the earlier revision's boards.

What is going on here?
Has Sega made further changes for Step 2?

And what is meant by "Field 1" and "Field 2" in the annotated board schematics?
Odd and even lines or what is meant here?



Thanks for any info!

I am asking just out of interest.

[gm_matthew]

The video board clock speed (50 MHz for step 1.x, 66 MHz for step 1.5/2.x*) applies to all of the ASICs and memory chips except for Jupiter, which is the ASIC that combines the 2D layers from the tilegen chip and the frame buffer to create the final video output; this probably runs at either 16 MHz or 32 MHz and is driven by the same clock crystal that drives the tilegen chip.

The main upgrade that the step 2.x video board provides is an extra pixel processor (Earth) and two more texture mapping units (Mars). These work very similarly to two Voodoo2 cards in SLI in that one Earth ASIC renders odd numbered lines and the other renders even numbered lines. This theoretically doubles the rendering fillrates compared to the step 1.5 video board.

Another thing added to the step 2.x video board is a proprietary DMA device (used for copying data from main RAM), since the NCR 53C810 SCSI device that fulfilled this purpose on step 1.x has been removed from step 2.x.

* To be really pedantic, the step 2.x video board is clocked at 66.666 MHz.

[Gjay]

Wow, thanks for clearing that up!
I was wondering about that for a long time...

So pixel- and texture-fillrates have been upgraded for Step 2.x.
What about geometry/vertices?


BTW, is the CPU-board found in Sega's Model 3 also part of the original Real 3D-pro image generator, or was the CPU-board completely composed by Sega?
I know the sound components were, but what about the CPU, RAM and the rest?

[gm_matthew]

Pretty sure the geometry processing ASIC (Venus) was not upgraded for step 2.x. From my own testing step 1.5 games have similar polygon counts to step 2.x games.

I can't be sure how much influence the original Real3D Pro-1000 had on the design of the Model 3 CPU board, but I'm pretty sure that Sega opted for a PowerPC CPU because that's what the Pro-1000 used. The address map for video board RAM regions is also the same. Probably the biggest difference is the 2D chip; the Model 3's tilemap generator chip is completely different to the "scroll layer" hardware on the Pro-1000, although it still communicates with the video board via the Jupiter ASIC.

[Gjay]

I don't see much of a difference between Step 1/1.5 and Step 2.x in general.
L.A. machine guns maybe looks somewhat ahead, but it seems Model 3 was nowhere near used to full capacity, especially Step 2.x.
(I mean where did all the extra fillrate go to? And what about that huge amount of available memory...)

I don't even see much of a difference in visual quality between Scud Race and Daytona 2, for example, even if there were about 1.5 years of time in between and the former ran on 1.5, the latter on Step 2. If I did not know this and someone told me it was the other way around, I would also believe that TBH.


So it is venus that is doing the geometry work... good to know.


But what I still am confused about:
If we talk abbout the Real3D pro 1000 "image generator" as a system as whole, does that include the CPU board or not?

I am asking because of this:
"The REAL 3D PRO-1000 interfaces with host computational devices by way
of an industry standard high-speed interface, allowing the user to select a host CPU ranging
from a PC to a high performance workstation to control real-time operation."

"The PRO-1000 is controlled by a user-supplied personal
computer or workstation that acts as host for control programs and as a link to the rest of the
simulation system. The PRO–1000 system off-loads intensive floating point computations
and all “out-the-window” image management tasks from the host computer to achieve a
sustained display update rate of 30 or 60 frames per second."

So... is this host computer the CPU-board in Model 3's case?

[Bart]

The Model 3 arcade board stack is similar to the Pro-1000 image generator as far as we can tell. The image generator had a PowerPC CPU as well that ran firmware. It was connected to a Windows NT workstation via a SCSI cable (hence why Step 1.x have a SCSI controller, which is used only for its DMA capabilities). Windows NT applications on the workstation constructed scene graphs and transferred them, along with polygon and texture data, to the Pro-1000, which then rendered them.

We haven't yet reverse engineered the firmware but as far as we can tell, the on-board PowerPC probably isn't doing much and is likely copying data that comes over the SCSI connection directly to culling RAM, ping pong RAM, and polygon RAM. Obviously there are some additional differences: Pro-1000 does not have a tile generator, does not have VROM (and hence, has more RAM for model storage), and does not have audio hardware. It has some sort of scroll layer, probably a simple frame buffer. There is also a serial port (which seems to be present on Model 3 boards as well, although no connector is exposed) that the firmware can communicate on and provide debug functionality (we can see lots of text strings embedded in the firmware ROM that correspond to this functionality and the manuals mention it).

[Gjay]

Thanks a lot for explaining... makes me unerstand the hardware a little more and I finally know what the SCSI-interface was used for.

Model 3 is just such an obscure piece of hardware, makes it even more interesting.
Some say it's because it was Military-grade hardware, so not a lot of info got into the public... and of course a company would keep it's intellectual property a secret.

But it's almost 30 years now (can you imagine?)...
So perhaps we should give Lockheed-Martin a call and ask nicely if they could please release all the relevant information about the Real3D pro. : D


How did you guys get involved into Model 3 emulation, if I may ask?

Do you just love the systems and it's games as I do?

[Bart]

According to Wikipedia, Intel purchased all of Real3D's IP and then 3Dfx purchased it from them. Then Nvidia and ATI fought over the corpse. There's nothing classified about Pro-1000 -- it was sold for civilian use cases such as flight simulators and some units were even sold to city planners (Philadelphia received one, I was told). Here's the craziest part: there is not one single photo of a Real3D Pro-1000 online outside of the marketing brochures. Not one single unit seems to exist anymore in the wild. I suspect those sold to the military were destroyed after being decommissioned and the civilian units were probably all junked as well because no one knew what they were (unlike surplus PCs, these would have been useless to people at the time).

I wish I had tried to track one down back in 2002-2003 when I first started looking at Model 3. The thought didn't even cross my mind. Now it's too late. I've contacted some people who worked at Real3D. No one has seen one since the 90's.

I got into emulation in high school. Wrote a Genesis emulator and then got curious about Model 2 when Richard Mitton released the first ever Model 2 emulator (it didn't emulate any of the 3D graphics but did display 2D tilemap layers). Model 2 was a beast. I looked at Model 3 next, which was easier to tackle (much simpler architecture), but didn't make much progress initially until I was introduced to Stefano Teso (who helped me figure out how to properly load the CROMs) and, later, Ville Linde, who was first to get 3D graphics up and running. We combined our efforts to create the early versions of Supermodel. Stefano and I got too busy with school and Ville then ported what we had to MAME and ran with it a bit further. I resurrected Supermodel in late 2010 because I was getting frustrated with my PhD and wanted an escape hatch to software engineering. I decided to brush up on my coding skills by working on something fun.

[Ian]

The Model 3 arcade board stack is similar to the Pro-1000 image generator as far as we can tell. The image generator had a PowerPC CPU as well that ran firmware. It was connected to a Windows NT workstation via a SCSI cable (hence why Step 1.x have a SCSI controller, which is used only for its DMA capabilities). Windows NT applications on the workstation constructed scene graphs and transferred them, along with polygon and texture data, to the Pro-1000, which then rendered them.

We haven't yet reverse engineered the firmware but as far as we can tell, the on-board PowerPC probably isn't doing much and is likely copying data that comes over the SCSI connection directly to culling RAM, ping pong RAM, and polygon RAM. Obviously there are some additional differences: Pro-1000 does not have a tile generator, does not have VROM (and hence, has more RAM for model storage), and does not have audio hardware. It has some sort of scroll layer, probably a simple frame buffer. There is also a serial port (which seems to be present on Model 3 boards as well, although no connector is exposed) that the firmware can communicate on and provide debug functionality (we can see lots of text strings embedded in the firmware ROM that correspond to this functionality and the manuals mention it).

Yeah the real3d pro-1000 was basically a standalone computer that received drawing commands from a SCSI interface. You could hook it up to anything really that would send it commands via SCSI, so linux, windows nt 4 or 3.5 etc. As far as I can tell the real 3d chip in the model3 and the stand alone image generators were almost identical. They've swapped some pieces of hardware around, ie the tilegen, sound, input etc but it's more or less the same. The real3d pro-1000 had gamma correction functions as well which is missing from the model 3.

The real3d pro-1000 has an SDK, which has a whole bunch of games / demos with, that run on windows. I tried once writing an emulator for this, which basically consists of an ASPI emulator. ASPI was a dll that windows used to talk to the SCSI hardware. I think modern interfaces are a bit different, but anyway, by creating a dummy dll and emulating the scsi I got a very basic emulator working. It was enough to get games to boot but not enough for them to run. I could never figure out exactly what the hardware should have been sending back. Basically games would get stuck in some sort of loop waiting for some specific data to arrive which I never sent.

If such hardware still existed, it would be super easy to log the commands and figure out exactly what should be returned by the hardware, and thus write an emulator for the actual real3d pro-1000. But sadly I don't think it'll ever happen. I've also never heard of anyone owning a stand alone board. Apparently they were very expensive, in the region of 20-30 grand which was a lot in the 90s. So there probably weren't even a huge amount of these made. Probably only found at universities or military etc.