3dfx Rampage ;)

[MistaPi]

okay here it is, but you didnt get this from me:

http://www.newgpu.com/3dfx/3dfximm2kdoc.ppt

F*ckers! I was saving that.

Been sitting on that since April 2000.

I have not seen this document myself (dead link), but I have heard that according to this document Rampage would have performance free AA.
This is supposedly from the document:

T-buffer based AA
-AA With 2 textures and Anisotropic with no performance hit.
-AA With 4 textures with no performance hit

I was under the impression that Rampage would be a standard IMR with multi-sample AA. Any comment on this?

 

[Laa-Yosh]

AFAIK Rampage had a flexible rasterizer with TMU's seperated from the pixel pipes. It was able to allocate the texture units for 2 or 1 pipe when multitexturing (that's 2 or 4 TMUs per pipe). The "idle" pipes then performed the multisampling, so for example with 2 textures you got 2x AA for free.

Not sure if I got it right though...

 

[no_way]

Nah ... Rampage utilized klonoatron flow through protomatter as most Steltek devices did.

 

[Xmas]

It's hard to tell what really is for free then. I think in this case AA comes at quite a high cost, but if you have AA enabled, it doesn't matter whether you use 1 or four textures.

Actually with this definition NVidia/ATI multisampling is (almost) completely free. Well, it is - as long as you don't look at the bandwidth.

 

[Randell]

ughh that reminds me, I need to get a Kyro II

was the Kyro II SE ever availible in retail?
our distributers doesn't carry it

As far as I know, no, you probably have to special-order it from Hercules or VideoLogic.

*starts bugging Guillemont and VideoLogic right away*

8)

I hope thats just to cpmplete a collection I would have thought there was no room for the KyroIISE with Herc producing Radeon9000's as their lower end card.

 

[Tagrineth]

AFAIK Rampage had a flexible rasterizer with TMU's seperated from the pixel pipes. It was able to allocate the texture units for 2 or 1 pipe when multitexturing (that's 2 or 4 TMUs per pipe). The "idle" pipes then performed the multisampling, so for example with 2 textures you got 2x AA for free.

Not sure if I got it right though...

It's something like that, yeah.

By the way, what Bonehed meant by 'no gates' isn't that the logic is entirely gate-free, but that it's kinda like an IKOS box - the transistors aren't really arranged specifically for one route, they are microcode-programmable so the gates are kinda virtualised. They're still there, just not fixed in place like normal. And yes, the rasteriser IS arranged in a logical fashion, not just one big array of transistors - just the gates aren't fixed in place... rather than no gates whatsoever

 

[RussSchultz]

By the way, what Bonehed meant by 'no gates' isn't that the logic is entirely gate-free, but that it's kinda like an IKOS box - the transistors aren't really arranged specifically for one route, they are microcode-programmable so the gates are kinda virtualised. They're still there, just not fixed in place like normal. And yes, the rasteriser IS arranged in a logical fashion, not just one big array of transistors - just the gates aren't fixed in place... rather than no gates whatsoever

Interesting. I wouldn't have guessed that an embedded fpga like product could produce competetive performance compared with programmable units with hardwired math functions.

 

[Tagrineth]

By the way, what Bonehed meant by 'no gates' isn't that the logic is entirely gate-free, but that it's kinda like an IKOS box - the transistors aren't really arranged specifically for one route, they are microcode-programmable so the gates are kinda virtualised. They're still there, just not fixed in place like normal. And yes, the rasteriser IS arranged in a logical fashion, not just one big array of transistors - just the gates aren't fixed in place... rather than no gates whatsoever

Interesting. I wouldn't have guessed that an embedded fpga like product could produce competetive performance compared with programmable units with hardwired math functions.

It isn't competely generic like IKOS... the transistors are actually arranged in a basic logical pattern, just without fixed gates... so they might be rearranged in the future for greater performance, and maybe a few extra features in the future. This way Rampage could probably be microcode-reprogrammed to support some levels of DX9 functionality (of course, not FULL but SOME), and such.

 

[darkblu]

By the way, what Bonehed meant by 'no gates' isn't that the logic is entirely gate-free, but that it's kinda like an IKOS box - the transistors aren't really arranged specifically for one route, they are microcode-programmable so the gates are kinda virtualised. They're still there, just not fixed in place like normal. And yes, the rasteriser IS arranged in a logical fashion, not just one big array of transistors - just the gates aren't fixed in place... rather than no gates whatsoever

Interesting. I wouldn't have guessed that an embedded fpga like product could produce competetive performance compared with programmable units with hardwired math functions.

It isn't competely generic like IKOS... the transistors are actually arranged in a basic logical pattern, just without fixed gates... so they might be rearranged in the future for greater performance, and maybe a few extra features in the future. This way Rampage could probably be microcode-reprogrammed to support some levels of DX9 functionality (of course, not FULL but SOME), and such.

tag, you don't need an fpga-like design in order to have microcode. as russ said, gate-level configurable hw can be nowhere as fast as a contemporary hardwired IC. i.e. while it's possible that the rampage was a microcode-centirc chip, it's _highly_ unlikely it had any considerable perts of it implemented as fpga.

 

[Tagrineth]

Well, whatever, all I know is, the performance is there, and it is microprogrammable at least to SOME extent...

 

[Sage]

yes, i was informed of the exact same thing as Tag. fully microcode programmable. i dont know how just how fast they were expecting it to go tho.


as for my Voodoo5, i chipped a ****tiny**** piece off of one of the chips and i looked very carefully and couldnt see any traces around it or anything that looked like a copper layer in it, so i went ahead and spent a few months ordering stuff / waiting for it to come in / cutting it up etc just to find out i had in fact killed the poor thing. when i put it in my system the thing beeps like there is no gfx card installed. does this indicate a killed master chip, or perhaps there is something else wrong (that i might be able to fix )? also, if you find out how to get a hold of a Kyro II SE, tell me!!!! i want one!!!! as for my V5, i just replaced it with a XFX Ti4200 that OC'd to 300/620 and i have yet to do the voltmod and put one xtra cooling (well i put AS alumina on it and some ramsinks on right out of the box)...... im thinking myabe a small peltier and a 1U copper CPU HS with a Delta 37cfm fan? oh yeah, and as soon as i get more AS alumina epoxy in im goinf to put a nice thick coat on the back of the card just behind the GPU to raise the surface area up slightly above the components and then slap on a mobo HS i poped off a display mobo at Frys... hey, its not like they should care, they put a giant wood screw thru the things to hold it down to the display. i dont think they'll have much use for a HS after that 8)

 

[Tagrineth]

yes, i was informed of the exact same thing as Tag. fully microcode programmable. i dont know how just how fast they were expecting it to go tho.

Message for you sir!

 

[Sage]

thanks Tagrineth, you're a doll
isn't she a doll, guys? (better agree or it'll hurt in the morning )

 

[Fox5]

Hey, I know this is old, but someone posted a link and I just had to jump in.

1. I don't think anyone was conspiring against 3dfx to make them fall to nvidia, but I do think some of the management basically cut the best deal for themselves.
2. If rampage was so good, I can't see nvidia not using it. If it was just slightly better, I could see them using their own stuff for pride reasons.(or if rampage was like twice the price)
3. I thought the initial specs 3dfx gave out about the rampage put the single chip version at the voodoo5 performance level, the dual chip at voodoo 5 6000, and a quad chip at twice that. So the mainstream part sounds more like a competitor to geforce 3, though I suppose maybe performance was underestimated and it could get that extra 50% to reach a geforce 4.
4. Didn't rampage have 128 tap af?
5. M-buffer?
6. Sage, how much did it hurt?

 

[Ailuros]

Vaporware season again!!!

If rampage was so good, I can't see nvidia not using it. If it was just slightly better, I could see them using their own stuff for pride reasons.

And who says that none of the technology hasn't been used since NV25? It should make more sense to keep the existing roadmap/research/development on track (whatever that means past NV2x heh) and incorporate those parts of the technology portofolio that can be implemented.

Examples would be amongst others, the 2D/VPE of Spectre, the "filter on scanout" trick, the infamous Rampage ROP trick etc etc.

I thought the initial specs 3dfx gave out about the rampage put the single chip version at the voodoo5 performance level, the dual chip at voodoo 5 6000, and a quad chip at twice that. So the mainstream part sounds more like a competitor to geforce 3, though I suppose maybe performance was underestimated and it could get that extra 50% to reach a geforce 4.

I don't recall them announcing any specs officially at all. Just some technology presentations at developer conventions, which I'm not sure anymore if they were publicly known back then.

Again for the last tired time:

value = single chip rampage/no sage estimated at ~150$
mainstream = single chip rampage + sage estimated at ~250$
high end = dual chip rampage + sage estimated at ~500$

Estimates would place the NV20 lower than high end yet higher than mainstream.

Didn't rampage have 128 tap af?

ROFL depends what exactly one means with anisotropic filtering exactly.

M-buffer?

Multisampling buffer; support for 2x and 4x sparse Multisampling and of course all T-buffer effects.

 

[Arun]

See what happens when you post a link to an old thread, just as a mere and simple joke
BTW, the "Why did the R400 get cancelled?" thread on the nV News rumor forums contains some of the worst anti-3DFX BS I've seen in years. "No, SAGE never existed, and Rampage never had more power than a mere Voodoo 5500".

Just wait till Tag sees that


Uttar

 

[Ailuros]

ROFL sure you criminal you, I'm currently looking for tar and feathers to publicly lynch you

Jokes aside, Fox5 had perfectly reasonable questions IMHO and I tried to answer some of them according to how I saw/see the whole story from my perspective.

 

[KimB]

2. If rampage was so good, I can't see nvidia not using it. If it was just slightly better, I could see them using their own stuff for pride reasons.(or if rampage was like twice the price)

Well, by the time nVidia actually acquired parts of 3dfx, the design was old. While many of the ideas that went into that architecture may have been good ones, the architecture itself was out of date. So, the engineers that nVidia gained were put to work on a new architecture, and the current designs were re-evaluated in the context of the newly-acquired architectural ideas.

 

[Arun]

BTW, 3DFX fans may take offense at this (although I still personally love the Rampage design from several POVs so don't take it personally ) but...
Is it just me or is the description of Rampage's AF pretty darn near to what XGI is doing today in their Volari (TM) series?


Uttar

P.S.: Please note that this sentence shows just how sad XGI's situation is; for any other IHV, it'd have been to their fanboys I'd have had to excuse myself, not Rampage's, because a 3-4 years old method is not quite impressive when it comes to AF. But in the case of XGI...

 

[Fox5]

Well, just searched the internet and found this at http://www.guru3d.com/arc18.shtml
3dfx Rampage Specs Leaked? - Xavier John @ 14:13

Entry Level Rampage Configuration:
2 way SLI (1 Rampage chip and 1 T&L unit)
32-512 MB memory support (200MHz DDR)
6.4GB/sec memory bandwidth
4 pixels/clock 800Mpixel fill rate (core clock at 200MHZ)
75 million triangles/polygons per second
AGP4x,
0.18 micron
32 bit rendering
Full DirectX 8.0 Features
Advanced T-buffer, Hardware T&L, Curved Surfaces, Full Pixel Shader, Quad texturing Photorealisitic rendering, Photoshop effects H2 2000
(ECTS at September, available in December)

Middle Level Rampage Configuration:
4 way SLI (2 Rampage chips and 2 T&L units)
32-512 MB memory support (200MHz DDR)
12.8GB/sec memory bandwidth
1.6Gigapixel fill rate (core clock at 200MHZ)
150 million triangles/polygons per second

High End Level Rampage Configuration:
8 way SLI (4 Rampage chips and 4 T&L units)
32-512 MB memory support (200MHz DDR)
25.6GB/sec memory bandwidth
3.2Gigapixel fill rate (core clock at 200MHZ)
300 million triangles/polygons per second


Based on these, wouldn't the midrange and highrange match up to some of the geforce 4's?

Also found this
http://www.vnroundup.com/index.php?item=articles/3dfx

The only downside would have been that MS takes the same texel coordinates and jitters them, thus there's no texture clean up. 3dfx implemented the advanced anisotropic filtering to remedy this once and for all. It could do up to 128-tap anisotropic filtering. If quad texturing was used, Rampage would provide for (performance-hit) free FSAA. When dual texturing was used with anisotropic filtering, Rampage again would have provided (performance-hit) free FSAA.