At Computex we saw something very interesting - both Intel and SiS based motherboards featuring dual graphics slots, AGP and PCI Express x16 to be specific. What makes these dual slot motherboards so interesting is that supposedly on VIA's chipsets actually featured both AGP and PCI Express x16 interfaces. So were Intel and SiS pressured into following in VIA's footsteps? We hit the show floor at Computex to try and find out.

Gigabyte had a board on display when we did our Early Bird Computex coverage that used Intel's 915 chipset and featured both slots, but when we went back to visit their booth the motherboard had mysteriously disappeared. In fact, when we asked Gigabyte about it they denied all knowledge of the board. Curious.

Click to Enlarge
Gigabyte's disappearing 915 board

ECS also had a board on display with similar properties, but unlike Gigabyte they were not afraid to show it. They informed all who asked that there was a performance hit when using the AGP slot on their motherboard (between 5 - 20%), but they didn't really explain how it was done. ECS also mentioned that you could populate both AGP and PCI Express x16 slots at the same time, but also, they didn't really explain how it worked.

Further inspection revealed that on both the Gigabyte and ECS boards there were no bridge chips or ASICs that would explain how they acheived adding AGP support to chipsets that supposedly don't support it. So were we just lied to? Does Intel actually support AGP and are they not telling anyone about it? As nice as that would be to believe, it's wrong.

Instead, what the manufacturers have done is gotten a little creative with their AGP interfaces and are actually running the slot off of the PCI bus, stemming from the South Bridge. Remember that the AGP interface is only 32-bits wide, the same width as the PCI bus and you can actually run an AGP card in PCI mode, giving up all of the benefits of AGP obviously. Remember that the PCI bus runs at 33MHz, offering only 133MB/s of bandwidth vs. AGP 8X's 2.1GB/s of bandwidth to the card. The incredible drop in bandwidth explains the performance hit, and although we haven't tested any of these solutions we'd say that running a fast AGP card on the PCI bus is probably not the best idea for performance. It was a valiant effort by these motherboard companies, but if you're interested in a high performance system then this isn't a solution. You'll either have to turn to VIA or bite the bullet and buy a PCI Express x16 card when they're available.

We believe that the SiS board we saw was most likely using a similar technique to achieve both AGP and PCI Express x16 support. This would also explain why the AGP slot on these boards was located far away from the North Bridge, which would normally be a routing/trace length nightmare.

Comments Locked


View All Comments

  • KristopherKubicki - Saturday, June 12, 2004 - link

    The AGP port has always been a hack on the PCI bus anyway - electrically it isnt that much different so this *does* make sense.

  • KalTorak - Monday, June 7, 2004 - link

    Derek - an x16 interface has sixteen diff pairs going each way, for a total of 64 signal lines. 4GB/s each way, simultaneously.

    This should also resolve your dilemma about the x1 interface - one pair in each direction, four total wires.

    As for getting a PCI-E spec, see if Anand can just join PCI SIG. :)
  • Minot - Sunday, June 6, 2004 - link

    Did Asrock do the same thing getting an AGP port on an Intel 865GV chipset motherboard? Their P4I65GV has what they call A.G.I. (Asrock Graphics Interface). It would make sense to me that this is also an implementation using the PCI bus interface. Or did they do something entirely different to get true AGP on that chipset?
  • MIDIman - Saturday, June 5, 2004 - link

    Hey - its a digital camera right? You should've just showed the Gigabyte people the picture that you took earlier! "Oh, that's clearly been doctored with Photoshop..."
  • Gandalf90125 - Friday, June 4, 2004 - link

    " can actually run an AGP card in PCI mode, giving up all of the benefits of AGP obviously."

    Can anybody say "3dfx"?
  • Bozo Galora - Friday, June 4, 2004 - link

    Much better more informative series of articles lately. Thumbs up.

    As far as the AGP pipeline, I think that it's the reason for most of the probs today with freezing, reboots and dumping to desktop. In the ATI smartgart tab you might find the AGP set to "0" (not 4X or 8X) by the self testing proceedure. What this means is that the AGP bus is shut off to the supposedly contiguous (gart) system memory address tables set in RAM. With a 128MB card probably, and a 256MB card definitely, it never has to offload work to the memory anyway, it does all work internally. People misrepresent this as no AGP, but if the card is set to 8X in BIOS, the cards mode is AGP VOLTAGE and IS much faster than AGP MODE (no busing involved), until it has to offload something, and cant (slowdown). We DO need a PCI high end videocard, because AGP 4X/8X is a failure. Remember the loop??? I want a NO AGP bus/gart table 1GB vram superfast memory PCI vidcard. Who cares about the bus.

    And heres your 89 page AGP 8X nightmare.

    I just bought a 9800pro and its a non functioning piece of crap. So now I have to go thru the endless driver in/out shuffle. I am really getting tired of this. The fact yours may work perfectly is irrelevant - for the average Joe, its a hopeless situation.

    Why the hell cant I just put in a card, load the drivers and have it work 100% - is that asking too much?? If a game or app doesnt work with a state of the art card they should not be allowed to sell it.

    What a business.

  • DerekWilson - Friday, June 4, 2004 - link

    Meh ... PCI Express nominclature (or maybe its explaination) could use some work ... maybe we'll do some in depth analysis for our first x16 vid card reviews ...

    The problem this: PCI Express for Graphics, which is x16, is defined (as I understand it) as 8 dedicated up and 8 dedicated down lines. You could call it full duplex with 8GB/s, or 4GB/s in each direction or whatever ... But does PCI Express x1 work the same way? It wouldn't be really possible since you'd have only one direction of communication (which wouldn't work for anything). So why don't they call the base spec x2 (since its got 250MB/s in each direction)? or why isn't the graphics spec called x8 (if each pair of up and down lines is called a connection as in x1)?

    I've gotta check with Anand to see if he's been able to purchase or aquire the PCI Express specification yet ...
  • Filibuster - Friday, June 4, 2004 - link

    >maximum of 500MB/s (8GB/s / 16)

    For full duplex, yes. Only 250MB/s in one direction, and almost all video traffic goes one way.
    Also, 133MB/s for standard PCI does not account for it's overhead either, so comparing the two theoretical speeds is probably best overall.
  • dvinnen - Friday, June 4, 2004 - link

    Seems like it would be easier to drop a nVidia-like bridge chip into the mother board it's self. I wouldn't think they need much tracing.
  • DerekWilson - Friday, June 4, 2004 - link

    oh .. and if these guys wanted to do build something with less of a performance impact, they could run the AGP card in 2x mode, drop in a buffer to build up 64bit packets of data (rather than the 32bits on the interface) and hook it up to a 64bit 66MHz PCI bus.

    That'd basically quadruple the bandwidth they're looking at right now with a glorified PCI slot.

    Of course, maybe my idea is harder to impliment than I'm thinking ... but engineers are engineers for a reason right? Crack the whip!

Log in

Don't have an account? Sign up now