Nappe1 does it again: OY!

[Nappe1]

Pyramid3D TR25203 (now known as VLSI VS25203) Hardware Reference Guide: http://www.vlsi.fi/datasheets/vs203_103.pdf

it's pretty hefty document (250 pages) so I don't expect good comments before tomorrow, but some highlights from this pice of Hardware demoed at DX6 Meltdown and at Assembly '97:

- Programmable Geometry Processor:
-- 3-issue VLIW processor
-- program lenght 16384 words
-- supports 32 bit fixed point and 32 bit integer types
-- supports branches and sub-routines.
-- data can be addressed from PCI, external memory or from other units inside the GPU.

- Primitive Processor
-- supports Z-Scale
-- supports a lots of things that I don't understand

- Programmable Pixel Shader
-- Z-buffer based pixel kill
-- 3 temp registers
-- program lenght 32 words
-- again, has much stuff that goes over my head.

Generally, it does all rendering on 32 bit and dithers the result, if needed. Geometry Processor seems to be much more capable than what nowadays "vertex shaders" are, but I can't be sure about that because most of that stuff goes a waaaayy over my knowledge. VLSI has been working on with Video Unit stuff. I don't recal that original Pyramid3D would have been capable all of that described on that part, but otherwise, it's definately the same chip.

 

[fresh]

I don't know if it's the same thing, but we actually had a pre-production Pyramid3D waaaaaay back. It was all wired together by hand, and of course - it didn't work

That doc is useless to anyone unless you wan't to write drivers.

 

[breez]

That's pretty interesting information I think. They were so much ahead of time. *sigh*

 

[Nappe1]

That's pretty interesting information I think. They were so much ahead of time. *sigh*

yep the point is that there's a lot of ppl saying now "wow! where we would be now if Rampage would have come out on time??" and I want to show that the industry is full of Rampage kind of stories. IF Pyramid3D would have been succeeded, we definately would have had programmability much earlier. IF Verite V4400 or Glaze3D would have succeeded, we MOST LIKELY would now have much more eDRAM based sollutions than we have now.... the list is endless.

anyhow, P3D's tech has some interesting points. For example, if I understood that document right, it's Geometry Processor was capable doing things like Displacement Mapping. (I might be wrong but, it seems to be pretty capable using any data as input.)

Of course it wouldn't be very fast doing it. but still nice feature.

is it that just me or does Pyramid3D have dedicated HW for Bump Mapping? IMO, it looks like that it does BumpMapping fully in pixel shader...

 

[Gollum]

The Pyramid3D would have been a very interestig product with great capabilities, the question however is, would it have performed decently enough compared to the competition? Some people in the know have said that the performance would hav been kind of lacking, so while beeing a very advanced piece of hardware technically, maybe it wouldn't have fast enough to make any of its features really usefull (just like one can argue that even Gf3 or R8500 might be too slow in more advanced shading operations to make extensive use of them and keep performance at an acceptable level). Still, they were way ahead of their time in some aspects!

On another note, I don't get the "if only Rampage would have made it to market" whining, PS and VS 1.0 were based off Rampage and from what I read about those specifications, they basically were much more limited implementations of the shaders concept what what ATI and Nvidia developed. It would have had a couple other interesting features, but overall it wouldn't have done much to advance 3D graphics as some people seem to think in their nostalgia flashs, it was way too late for that already (again)...

 

[Nappe1]

The Pyramid3D would have been a very interestig product with great capabilities, the question however is, would it have performed decently enough compared to the competition? Some people in the know have said that the performance would hav been kind of lacking, so while beeing a very advanced piece of hardware technically, maybe it wouldn't have fast enough to make any of its features really usefull (just like one can argue that even Gf3 or R8500 might be too slow in more advanced shading operations to make extensive use of them and keep performance at an acceptable level). Still, they were way ahead of their time in some aspects!

and maybe one of the key things is the combination of very highly advanced features and not just the lack of speed but more like support on APIs... back in DX6 times it was pretty useless even have something like 3DLabs' GLiNT Geometry Processor or Pyramid3D's Geometry Processor, just because there really wasn't way to support it easily. And when DX7 was out, it was absilutely too late to it's speed.

So most likely one of it's problems would have been the API support. not earlier than DX8 from DirectX'es can reveal majority of P3D's features. So, maybe one of their problems was that they truly wanted to beat the competion with giant leap. if we look at the nVidia, 3dfx, rendition and ATi on summer 1997, how many of these would have been capable keeping up on race? So, my point is that maybe their feature set was too advanced?

and as an IMO to that last my own question: Rendition would have been closest one to keep up on programmability. and again, world would be pretty much different...


the real reason why I dig up this document, is that this is the closest FACT we can get what P3D was / would have been. I know there's more than few working samples of different revisions exists, but it's most likely that no one would dig up a one and test, what's the real performance of chip. As said in document. it has fill rate of 60 Mpixels (dual textured.) so it would propably have head to head against Riva 128 and Voodoo Graphics, but how it would have compete, will stay a mystery.

 

[ERP]

As said in document. it has fill rate of 60 Mpixels (dual textured.) so it would propably have head to head against Riva 128 and Voodoo Graphics, but how it would have compete, will stay a mystery.

I can only comment with second hand informaton here, since we never purchased a dev kit. However I do know of a couple of people who did and both had similar comments.

On paper the design looked great and the feature set was truly compelling, but the real world performance compared to available silicon was lacking.

 

[megadrive0088]

Ahhhhhh. This Pyramid3D was the hardware I wanted most, aside from the Lockheed R3D-100. It's one of the few chips back then that had a geometry engine. probably the only consumer chipset with geoemtry processing at the time. Pyamid3D's chip configuration, at least from the outside, reminds me of the R3D-100, the -100 had geometry procoessor, texture processor and graphics processor. though they were seperate. Pyramid3D had all its processors on one chip.


good find Nappe

 

[Simon F]

As said in document. it has fill rate of 60 Mpixels

Did it really? I'm sure it was demonstrated here and it seemed to me to be a bit slow.

 

[megadrive0088]

I soooooo wished that:

Lockheed Real3D
Tritech Pyramid3D
Videologic PowerVR

had been -completely- sucessful in the consumer-desktop space and thus game consoles (and arcades) as well. I never really liked 3Dfx, nVidia or ATI.

At least PowerVR had a degree of success in all three.

 

[Mariner]

Nappe1 wrote:

As said in document. it has fill rate of 60 Mpixels

Did it really? I'm sure it was demonstrated here and it seemed to me to be a bit slow.

Wasn't the problem that it would have been lacking somewhat in bandwidth compared to Voodoo/Voodoo2 et al? And, of course the deferred rendering architecture of the PowerVR offering of the time!

I used to have the various PDF presentations that were around but can't find them to check up. I remember they said a wide range of memory interfaces could be used.

 

[Nappe1]

it was lacking bandwidth. it had 64Bit bus while Voodoo 2 had 2x64Bit independent busses.

Pyramid3D has same bandwidth as TNT Vanta: 800 MB/s


and acctually there was some great news considering Bitboys in one finish magazine. But if anybode else does not tell those earlier, I'll delay that a bit so that there won't be "NV30 released and what you know, Bitboys show up again" -problem. (though this news and NV30 has nothing common, there's still possibility that someone takes it that way.)

 

[LeStoffer]

...and acctually there was some great news considering Bitboys in one finish magazine. But if anybode else does not tell those earlier, I'll delay that a bit so that there won't be "NV30 released and what you know, Bitboys show up again" -problem.

Just tell us about it now...

 

[Nappe1]

okay, here it comes... hopefully I am not doing a biiig mistake.
at the Talouselämä -magazine this years latest issue (#39), has a small article about Bitboys: (first the original quote in finnish and then translation.)

Bitboys pääsi Japaniin
..Bitboys toimii nyt 21 hengen voimin Noormarkussa ja Espoossa. Toimitusjohtaja Juha Taipaleen mukaan yhtiön liikevaihto nousee kuluvalla vuosinelänneksellä 700 000 euroon ja liikevoitto 150 000 euroon.

and the my own translation:

Bitboys goes to Japan market
...Bitboys works now with 21 employees at Noormarkku and at Espoo. Acording to CEO Juha Taipale, company's turnover raises up to 700,000 Euros and the Operating profit raises up to 150,000 Euros on this quater.

thanks to Nemesis77 from Muropaketti.com boards bringing this up. I also have to agree with him that first time I hear them talking about profit.

EDIT: fixed translation errors...

 

[Simon F]

it was lacking bandwidth. it had 64Bit bus while Voodoo 2 had 2x64Bit independent busses.

In the words of Monty Python's yorkshiremen, "Luxury". PCX1/2 had a 32 bit bus!

Still, having had a quick glance at the manual, it seemed a very interesting chip.

 

[Johnathan256]

At least PowerVR had a degree of success in all three.

I am keeping my fingers crossed about the PowerVR Series 5 at the beginning of next year. Could show us future of deferred rendering(or tiling at least) I think PowerVR has some more damage to do.

 

[Nappe1]

it was lacking bandwidth. it had 64Bit bus while Voodoo 2 had 2x64Bit independent busses.

In the words of Monty Python's yorkshiremen, "Luxury". PCX1/2 had a 32 bit bus!

Still, having had a quick glance at the manual, it seemed a very interesting chip.

yep, they had.
but you have already noticed that PCX1 had only point sampling and afaik it could fetch only samples from one texture map per clock (am I right?). Pyramid3D has 2 TMUs capable fetching 4 samples from 2 texture maps per clock.

so maybe I should have said: "unbalanced" instead of "lacking the bandwidth". And fact is that so far no one here has first hand knowledge of it's performance, but many has said that they have heard it was rather slow. Another point is that back on it's "dev kit days", there was only DX6 drivers available, which eventually didn't use more than fraction of the chip's capabilities. (for example Geometry Processor wasn't supportted at all.)

but, propably we will never know the exact information HOW MUCH it was lacking behind Voodoos (1 & 2), Verite chips, PCX 1 & 2, Riva 128 and Rage II.

I also find out that Bitboys and VLSI were doing chip design as co-operational work from the begining. Because their co-operation ended after Pyramid3D project, there propably many things went wrong.

 

[Sage]

On another note, I don't get the "if only Rampage would have made it to market" whining, PS and VS 1.0 were based off Rampage and from what I read about those specifications, they basically were much more limited implementations of the shaders concept what what ATI and Nvidia developed. It would have had a couple other interesting features, but overall it wouldn't have done much to advance 3D graphics as some people seem to think in their nostalgia flashs, it was way too late for that already (again)...

you are forgetting that the Rampage had updatable microcode. it is capable of many DX9 features with an updated microcode. but alas, who actually has the few pre-production chips? even if a private citizen or other company had them (ie anyone but nVidia hehe) who would be able to update the microcode? i'll bet any Rampage's / Sage's nVidia has are locked up in a hidden safe somewhere so that they might be used for reference if absolutely nessesary. too bad the 3dfx Rampage Project's results will probably never be seen by the world. But then, I could go on and on about Pyramid3D, V4400, Real3d-100, G800, etc....

 

[Nappe1]

you are forgetting that the Rampage had updatable microcode.

someone have not A) read the doc I linked or B) understood what he has been reading. The programmability of pixel pipeline at Pyramid3D was based on microcode. Maybe that's the reason why turned out to be so slow. The Geometry Processor was a real processing unit, but the Pixel Pipeline programmability was done with microcode. Still, the microcode isn't any magic wand that would be the key turning DX8 card to DX9, in the case of Pyramid3D, there was only 3 temp registers on pixel shader and you could not do anything to that on microcode. you just had live with that. (DX8 wants 6 temp registers, DX9 much more.) So, if you want to make GFX card with upgrade possibilities you are talking about, then you have to make part of the chip to next generation design allready.

look... Pyramid3D was designed in 1997. though, it's internal precision doesn't match to nowadays cards, it's Prog. Geometry Processor flexibility still surpasses R300 and NV30. afaik, it's close to support VS 3.0 (DX9), but I can't be 100% sure about that. Pixel Shader is near to PS 1.1 (DX8) and we are talking card that was introduced to compete against Voodoo Graphics. (Voodoo 1)

so if you want to put it your way, Pyramid3D was DX6 card able to support at least some new features of Direct 3D's all way up to DX9. just with updating code that goes to GPU.

 

[Simon F]

Still, having had a quick glance at the manual, it seemed a very interesting chip.

yep, they had.
but you have already noticed that PCX1 had only point sampling and afaik it could fetch only samples from one texture map per clock (am I right?).

I 'noticed' no such thing. PCX1's native mode was linear MIP mapping (i.e. blending samples from two Map levels as in Dungan's precursor to William's paper). It did have a tiny cache to reduce reads though. At the time it was felt that the 'texture aliasing in the distance' problem was the most important to address. Remember this chip was developed around the time of Doom which had rather nasty aliasing.

Pyramid3D has 2 TMUs capable fetching 4 samples from 2 texture maps per clock.

Do you mean 2x4 samples or 4 in total?

so maybe I should have said: "unbalanced" instead of "lacking the bandwidth". And fact is that so far no one here has first hand knowledge of it's performance, but many has said that they have heard it was rather slow.

Unless I'm mistaking it for one of its contemporaries I'm sure a demonstration was given here in Kings Langley.