3dfx/3dhq Revenge superchip.

[Tagrineth]

Ah well, wrt shared RAM, the new multi-chip tech CAN share the same memory, unlike good old SLI.

I don't know enough to answer most of your questions right now... sorry. One side note though, about OpenGL Guy's gates, gates, gates discussion, the theoretical Revenge was arranged a little differently than one would expect. As I said though, my own knowledge doesn't go deep enough to explain much.

Though... yes, there are some big discrepancies and contradictions. But, for one, think of the 3.2Gpps fill rate as an 'effective' one, not theoretical.

 

[Pete]

Obviously the specs are pie-in-the-sky, and the story surrounding the chip is so dramatic it sounds farcical, but I'm most interested in the poster's Rampage numbers. Are those for real, or did the poster lie about everything?

Aqueel tested in in Quake 3, and @ 1280x1024 with max gfx @ 8xAA it churned out 254FPS.

 

[arjan de lumens]

Though... yes, there are some big discrepancies and contradictions. But, for one, think of the 3.2Gpps fill rate as an 'effective' one, not theoretical.

"Effective"? Like the ~3x actual fillrate numbers that IMG used to be claiming, due to the high pixel rejection rate of their architectures? For reference, what would the "effective" fillrate of R350 be, given its HyperZ?

Hmmm - when I think of it, a tiler with two 500 MHz pipelines, 32 Z-test units, and "effective" fillrate = 3x actual fillrate, could actually match what these guys are claiming (including the transistor count and power dissipation), but would almost certainly infringe IMGTEC IP in a big way.

And as for shared memory between multiple chips: that would require an extremely high-bandwidth interconnect fabric between the chips to be useful - a bit like AMD Opterons, but with still more bandwidth, presumably adding to the cost of each chip.

 

[KimB]

Quick side note on the 64-bit z-buffer:

This just sounds like a stupid idea to me. We should go full 32-bit z before going for 64-bit. A 32-bit z-buffer should be enough for most everything. Of course, this would require separation of the z and stencil buffers, but I think it's about time for that to happen...one thing that might work is use part of a packed framebuffer type to include stencil.

 

[Tagrineth]

Pete: The Rampage frame rate is real. That's 8x multisampling with 16x AF.

Arjan de Lumens: Pretty sure they avoided IMG's patents... and yes, the RAM architecture is another of Revenge's 'hazy oddities'.

Chalnoth: Radeon R200 and VSA-100 already support full 32-bit Z, though for some inexplicable reason R300 doesn't. Of the "big three", only nVidia has yet to support 32-bit Z. However, I should note that a 24-bit W-buffer tends to be more useful than 32-bit Z (though again, curiously enough, R300 doesn't support that either).

Another general aside: Rampage, on top of that speed, supported 52-bit INT.

 

[OpenGL guy]

Pete: The Rampage frame rate is real. That's 8x multisampling with 16x AF.

Suuuuure it was. Note that the performance is higher than the CPU probably could supply.

Chalnoth: Radeon R200 and VSA-100 already support full 32-bit Z, though for some inexplicable reason R300 doesn't. Of the "big three", only nVidia has yet to support 32-bit Z. However, I should note that a 24-bit W-buffer tends to be more useful than 32-bit Z (though again, curiously enough, R300 doesn't support that either).

As long as we keep doing computations in 32-bit floats (read: source data from app.), 24-bit Z is plenty. 32-bit float has 24-bit mantissa... barely enough.

 

[Tagrineth]

Pete: The Rampage frame rate is real. That's 8x multisampling with 16x AF.

Suuuuure it was. Note that the performance is higher than the CPU probably could supply.

Was the top-end CPU of the time, I forgot what the exact clock was though. Was a well-over-2GHz P4, which I'm sure could supply enough for that frame rate.

 

[Testiculus Giganticus]

Hmm, i do think that this rampage nonsense should really stop.Sorry Tagrineth, but it is unbelievable that a 3 year old chip could sport that kind of framerate, no matter how revolutionary, and considering that it was an IMR, and rather traditional in many ways.Not to mention that even modern day cpus hardly provide enough power to reach those numbers.

 

[Geeforcer]

Yes, Tagrineth, whatever you say...

3dfx went down in December 2000. In the end of April 2001 ( 4 month later), the fastest CPU availible was P4-1.7GHz.

In Firing Squad's review, a P4-1.7@1.9 was able to achive a score of 254fps at 640x480 Normal.

To put this BS further in prospective:
http://firingsquad.gamers.com/hardware/ati_radeon_9800_pro/images/4aa8xaniso1280.gif

Wow, apperently Rampage 8xFSAA/16x Aniso was almost 2 times faster then 9800 Pro 4xFSAA/8xAniso... Can we please stop this nonesense once and for all?

 

[Dave Baumann]

Considering that with 4X MSAA enabled an entire Rampage chip would have effectively been one GF4 pipeline, but with twice the texture rate, I'd doubt those numbers!

 

[Tagrineth]

Testiculus Giganticus: Think a decent bit more powerful than one chip from that timeframe, then slap another one on the board and link them together with 99% efficiency.

Geeforcer: The Rampage board wasn't in a system made shortly after 3dfx went down, THANK YOU. Something closer to middle of last year.

Another thing to note is that Rampage could do AF effectively free with MSAA enabled.

DaveBaumann: And you know that as fact, set in stone? You think Rampage had no functions to improve MSAA rate? Especially when multitexturing is involved... One of Rampage's niftier features for MSAA speed (among other things) is, that while multitexturing, the additional pixel pipelines that wouldn't be used while looping back or TMU-combining could be used to fetch additional MS samples.


Well, anyway, that IS enough, more or less.

 

[Ailuros]

Wow, apperently Rampage 8xFSAA/16x Aniso was almost 2 times faster then 9800 Pro 4xFSAA/8xAniso... Can we please stop this nonesense once and for all?

Combine that with Dave's post above and add that there was never a higher than dual chip rampage/ 1 sage=Spectre planned due to cost restrictions. 500$ estimated back then for that with just 200MHz DDR.

MSAA on Spectre was NOT fillrate "free" (relatively speaking).

 

[Ailuros]

Another thing to note is that Rampage could do AF effectively free with MSAA enabled.

See also above. With dual texturing 2xRGMS + 32-tap on high end (ergo dual chip) Spectre, it came essentially with minimal performance cost.

Only Fear was to have true (still relative) free MSAA.

I'm pretty sure you incurably romantics know all better than folks like Dave (x2 hehehe), Reverend, Kristof and so many others around here.

Give it a rest. 3dfx is dead and burried since 2000.

***edit: by the way 8 textures/pass and 4 textures/clock

 

[Xmas]

Well, anyway, that IS enough, more or less.

No it isn't. Do the math.

254 fps @1280x1024 with 8xMS is 2,66 GSamples/s.
That is, with 100% efficiency and no transparent overdraw at all. With Q3 average depth complexity that would be >4 GSamples/s.
Rampage specs were far too low to reach that.

 

[Ostsol]

Tagrineth, think about it: NVidia owns 3dfx and all intellectual property and techologies that 3dfx had, including any research 3dfx was doing. NVidia doesn't have 8x FSAA that runs that fast, meaning these Revenge/Rampage numbers are BS.

 

[OpenGL guy]

Well, anyway, that IS enough, more or less.

No it isn't. Do the math.

254 fps @1280x1024 with 8xMS is 2,66 GSamples/s.

It's not as bad as that if you have compression. I don't know if Rampage had that feature.

 

[Fafner]

Tag,
did you see this yourself, or did aquoess provide you with the numbers?

Faf

 

[Geeforcer]

Sigh...

Let me re-iterate the claims made here by Tagrineth:

A Rampage board is capable of running (at 1280x1024) 8xMSAA/16xAF at 254 FPS.
A 6-month old Radeon 9700 Pro gets 280 fps with NO FSAA and NO AF.
A brand-new Radeon 9800 Pro gets 167 fps with 4xFSAA and 8xAF.

Rampage specks:
Fill rate: 800-1000MP/s (one chip), 1600-2000MP/s
Bandwidth: 6.4Gb/s (one chip), 12.8Gb/s (two chips)

Radeon 9800 pro specks:
Fillrate: 3040MP/s
Bandwidth: 20.48GB/s
HyperZIII, Color compression, crossbar memory controller and other features to maximize effective bandwidth utilization.

 

[Tagrineth]

Hmm, I just thought of something. I'll be back to this thread in a couple days. 'skooz.

 

[Xmas]

It's not as bad as that if you have compression. I don't know if Rampage had that feature.

Compression doesn't save the fill rate. Bandwidth may have been sufficient, but Rampage didn't have fill rate free multisampling, as was already pointed out here.
And even if it had, i'm sure you agree an 18 GSamples/s R350 with compression would beat it