hey guys,

I know, nice topic subject, but that was my question.

in theory, 2 5500 pci cards would equal a 6000, IF you could get it to work together, but that is probably not possible or is it??

Can drivers be "hacked" so that in some way the 6000 drivers could be used for the 5500x2?? (no ATI pun intended)

I would love to know if this is possible AT ALL?? in theory, could it be done?? what would be the most likely way to get it to work??

Will it never work due to there not being a bridge chip like on the 6000, or was that just so the agp port would identify all 4 chips on the same card??

It is possible, but this would mean a lot of driver re-writing.

No matter how, the biggest problem would be hardware limitations. A normal PCI bus with 33MHz already limits a single 5500 and now would additionally need to carry the traffic between both cards.

What the bridge on the V6k does is to generate a second PCI bus within the card itself to connect alls 4 VSA100 GPUs. It works with 66MHz.

So PCI-X with 66MHz is at least necessary to keep the performance up. But that's not all...

The VSA100 SLI design needs one master GPU with whom all further VSA100 communicate with. That's a lot of traffic between both cards, not to mention additional add-in cards like sound and network.

just a little brainstorm from my side:
- 66MHz PCI bus
- driver needs to declare the master card
- master card gets the image processing data
- master VSA100 needs to share data to 2nd, 3rd and 4th VSA100 (the bottleneck is connection between master VSA100 and 3rd / 4th VSA100)

Would be a nice project.
 

that sounds pretty logical to me :)

but would it be possible to do that with 1 pci and 1 agp?? that would then remove the bandwidth limit of the pci bus, wouldn't it??

I do have a dual P3 1.4Ghz with pci-x 66Mhz slots, even 2 133Mhz if I'm not mistaken, but I thought only the MAC 5500's could run in a 66Mhz slot?? Or is it the only one that can use the 66MNhz slot fully??

But in case of the driver rewriting.......it would be best to use the 6000 drivers wouldn't it?? since that already has that sort of "setup" if you can call it that?!

I mean, if this would work, then in theory, you could add as many 5500's as you want, as the VSA-100 limitation was max 32 chips per card, wasn't it??

Imagine a rig with 1 x agp and let's say 4 pci version, all on a 66Mhz or 133Mhz slot, that would be uber pownage :D VOODOO style!!  :o :o :o :o :o

ultima wrote on 13.12.12 at 18:51:10:

but would it be possible to do that with 1 pci and 1 agp?? that would then remove the bandwidth limit of the pci bus, wouldn't it??

Should be possible too, just look at the simFUSION 6000q :) The Voodoo 5 5500/6000 only work with AGP 1x, which means it is more or less PCI at 66MHz. So we would have more bandwith for the master card, the slave card would be limited to 33MHz.


ultima wrote on 13.12.12 at 18:51:10:

I do have a dual P3 1.4Ghz with pci-x 66Mhz slots, even 2 133Mhz if I'm not mistaken, but I thought only the MAC 5500's could run in a 66Mhz slot?? Or is it the only one that can use the 66MNhz slot fully??

Normal Voodoo 5 5500 PCI also work with 66MHz if put in a PCI-X slot. So two 5500 PCI in two PCI-X slots is the best configuration for this scenario :)


ultima wrote on 13.12.12 at 18:51:10:

But in case of the driver rewriting.......it would be best to use the 6000 drivers wouldn't it?? since that already has that sort of "setup" if you can call it that?! I mean, if this would work, then in theory, you could add as many 5500's as you want, as the VSA-100 limitation was max 32 chips per card, wasn't it??

Between 5500 and 6000 there isn't much difference within the drivers. The workload is just splitted to 4 GPUs instead of 2. Same when it comes to AAlchemy drivers. The initialisation routine is the first problem (the driver needs to decide which card is master and which one is slave) and the second one is the data routing on the normal PCI bus. That's at least what I think.

If theses problems are solved 32 GPUs might be possible in theory :)

But since I'm not a driver developer I can't tell if this is a realistic scenario.

Quote:

Should be possible too, just look at the simFUSION 6000q The Voodoo 5 5500/6000 only work with AGP 1x, which means it is more or less PCI at 66MHz. So we would have more bandwith for the master card, the slave card would be limited to 33MHz.

Actually, 5500 / 6000 run AGP 2x, so that would be 133Mhz.
So we need someone who can write drivers?? This would be a nice project indeed, this would allow the "small folk" to enjoy the power of a 6000 by combining 2 or more 5500's, that would be awesome in my opinion.

Maybe we can check with AmigaMerlin?? He writes drivers for voodoo5 cards, doesn't he??

ultima wrote on 14.12.12 at 08:49:00:

Actually, 5500 / 6000 run AGP 2x, so that would be 133Mhz.

Nope only 66MHz ;) This is because of the SLI Link between the VSA100. A 5500 does have a direct PCI bus between the two VSA100 (it works with 66MHz), the 6000 uses the bridge to generate a PCI bus between all four VSA100 (66MHz too).

The AAlchemy also uses several bridge chips to connect all  eight VSA100, but the bus only works with 33MHz internally and that's why it only works with 33MHz to the PCI-X slot of the mainboard. The card can't handle more bandwith.

A single VSA100 is capable of AGP2x speed and compatible to AGP4x voltages (but nothing more).

We are limited to 66MHz :(

AmigaMerlin or SFFT. But I don't know if they have enough know-how to write a new initialisation routine.

OK, I'll take your word on that :) you know more about that then me, so.... :)

As for AmgiaMerlin or SFFT, seeing as they have been writing drivers for the 5500's for quite some time now, I'd figure they would know how to do this as well, or at least pinpoint where in the drivers the initialisation route is determined.

Maybe it can be cross-referenced with the voodoo2 drivers, seeing as those would need a "master card" as well, right??

***EDIT***
I was reading up a bit and saw this post:
http://descenttaskforce.forumotion.com/t40-3dfx-voodoo5-6000-voodoo5-5000-world-owners-lists-research

stating:

Quote:

ASUS A7M266-D Rev.1.01 incompatible, Rev.A 3700

seeing as this is my mobo what I use for my 5500AGP, is there any chance this could be incompatible with this project to be?? Or is it just HinT chip incompatible??

Sure, maybe the Voodoo 2 drivers might be able to give a hint how to get such a thing working. But our driver developers should be able to tell us more about that.

As for that board, I think it's only HiNT bridge related. Not because of the voltages but maybe because of some timing issues.

that's good to hear, cause what I totally forgot, but saw again today when starting to test several cards, is that this board has 1 AGP slot AND 2 66Mhz PCI slots, so if in the future this project gets off the ground and we get it working, I can run 3 Voodoo5's in 1 system.

They'll have more then enough cpu power, as I am currently negotiating about a 2nd XP-M 3000+ :)

I thought about this a long time ago, but I've no idea how drivers works.
But can we imagine that the drivers only sees 2 5500.
It sends the first frame be calculated to the first 5500, the second to the second 5500, third to first 5500, etc. So no more need a master VSA100 which communicate with the 3 others VSA.

And to get the final imagine I don't know the method ; maybe a third card which take the pictures from the framebuffers of the two 5500...

in my opinion, if we can get it to work that 1 card is master and 1 is slave, then it shouldn't be too hard to define a 2nd slave either.

It should be in a setup like; the 1st chip on the AGP card will be the master and all other VSA-100 chips will reroute to that one, so .....

I have downloaded the source code drivers for the voodoo5, but those are linux I think :)
I have also downloaded the newest amigamerlin drivers for voodoo5 and the voodoo2 drivers, to see if I can spot any clues.

I am now very happy, since I traded 1 of my 5500 AGP's with someone for a 5500PCI, so now I have 2 PCI versions and a motherboard that can run both of them on 66Mhz pci buses. :)

Nice :)

Would be nice if someone could provide us with the W2k/XP driver source.

maybe we can contact AmigaMerlin or SFFT for that?? maybe nice to try on linux 1st?? since that is all open source and should be easier to find??

But can't we use the existing 5500 drivers?? start with the basics with the last WHQL drivers from 3dfx themselves??
Can't we "reverse engineer" those drivers so we get the source code? or am I talking BS now?

There was a guy on this forum that created his own drivers for V5. Maybe he can jump in?  ::)

do you know who it was??

ultima wrote on 07.02.13 at 16:09:40:

do you know who it was??

Geri:
http://www.falconfly.de/cgi-bin/yabb2/YaBB.pl?num=1268949568

3dfx's driver sourcecode is on v3info:

http://3dfxbios.stantoworld.co.uk/files/3dfx_source_code.rar

I sent Gery an email asking for some info along with this post, hoping he can shed some light on the subject and if he knows how such a thing could be accomplished.

I got a reply from Geri, I guess the idea was ok, but technically, it is an outright no-go:


Quote:

Hi Geri, I've followed your driver development a bit on the background, but is it possible for you to rewrite 3dfx drivers so that multiple 5500's can be used in the same system?? I have been discussing this on the board and the problem is the initiation route, which vsa-100 is master etc etc. http://www.falconfly.de/cgi-bin/yabb2/YaBB.pl?num=1355402726/0 Could you let me know if it is possible? kind regards Pim

GERI:
there is a problem when using multiple v5 5500 would occur. the driver would require to sync the rendered image manually. the problem is that copying the frame buffer, and pasting it together would be probably slower, than just rendering the image on one simple card. copying and pasting back to gpu on a v5 would clamp the speed to probably somewhat around 5 fps. v5 is dead.


Quote:

so in short, it would be possible, but the disred effect is not feasable? But how does this work with the voodoo 2 then? Is it really cause of the hardware SLI connector that that does work and the 5500 would have to do it through software and pci bus?

GERI:
yes, when v2-s connected with sli bus, they sync the frame buffer directly, without requiring software texture transfer.

the same happens beethwen vsa100 chips when they are electronically connected, like voodoo5.

to make this with software assistance, it would need
-at least agp 8x transfer speed to be effective
-and support for virtual frame buffers and rendertotexture caps, OR support for hardware assisted quick framebuffer to texture copying.

3dfx cards dont have any of these requietments. this is only possible with some early cad cards, and above radeon 8500 / geforce3.

copying the frame buffer from voodoo5 to texture, takes 0.04-0.1 second.
copying the texture out from voodoo5 to system memory takes additional 0.05-0.1 second.
copying the half texture back to another voodoo5 would take ~0.02-0.05 second.

copying the frame buffer out from the chip decrases the performance below 4-8 fps.
and then you didnt even rendered anything, you just synchronized the images.

making this is possible with newer graphics cards, becouse they have enough bandwidth to move the images. but not with 3dfx.

and there is another problem with this too. since 3dfx cards cant do hardware vertex transformation and lighting, the voodoo5 need a lot of cpu power to  process the vertices. the cpu just simply cant feed your voodoo5. and in this case, it would need to feed two of it.

to be short: voodoo5 is just a little improved voodoo1, that can render bigger textures in 32 bit, and have 2d functionalities. but its internal command set still looks back to somewhere around 1995. this is why voodoo5 is so slow.

Quote:

Well that really sucks, since I remember reading not to very long ago, cpu's were finally so fast, they could do software rendering as a voodoo2 or 3 could hardware wise do oh so many years ago. Reading your explenation, this would require something like a dualxeon based system, both at least core i7 based quadcores at 5Ghz, and  2 pci 5500's on a pci2pci-e connectors and then you'd might be able to emulate a 6000, right?? was a nice idea, can't the floppy like connector be used for it hardware wise??

GERI:
most modern software renderers, such as llvmpipe with mesa9.2 alreday perform around Radeon HD 5450/HD6450 on a moden 4 core i7 sysetm. (uncomparable faster than any voodoo).

also, you will still not be able to create a faster platform for two voodoo5 connection even if you put an 5 ghz cpu inside, since the communication is limited by the voodoo5 and the motherboard chipset itself. i dont think there is a possible way to connect two voodoo5 effectively.

How are actually V5 5500 and 6000 chips organised? Do they render 1/2 and 1/4 of the frame or do they render 2 and 4 frames in advance?
I see why it would be slow and impracticle to use V5 5500 if they do render one frame and sync, but what about if they do sequential frames?

The image gets "splitted" into rows.

1st VSA100 => uneven rows
2nd VSA100 => even rows

11111---------------11111
22222---------------22222
11111---------------11111
22222---------------22222
...
...
...

The V5 6000 is doing it the same way.

So it's some sort of SFR (split frame rendering). AFR (alternate frame rendering) would also need to transport the frames from one card to the other... bandwith problem as Geri already mentioned.