And you'll probably find the AGP bridge isn't either, which may be another
cause of slowdown.  Oh, and the XFree86 DRI stuff won't help at all either
- the mainline X tree is *old* in this regard.  If you're feeling brave,
check out the current XFree86 snapshots - ATI have been contributing a
*lot* of code, and I think they've got 2D acceleration for the ATI IGP
chipsets in there now (SWMBO has a laptop with one of these, and she's
asking me to do this...).  If you're *really* lucky, the latest X snapshot
will actually build!

I can't speak for the ATI support in XFree86.  But we have a bunch of Intel 865 boards here that have integrated i830 video chips.  Their 3D performance (for simple-ish stuff) is quite good using DRI/DRM.  Though some of the GL screensavers tend to crash X.  And with the i8x0 chipsets you're limited to 16bit colour if you want 3D.

This tells me that your app doesn't use GLX, so it's probably doing all
the rendering using the machine's CPU, RAM, $deity knows what else.  As
the Duron 900 is the slowest machine you've got.

I'm still boggled at anything over VNC being the fastest, though.  That's
just bizarre.

The biggest problem historically with graphics has always been the bandwidth between video memory and main memory.  Graphics are just plain big and eat the data bandwidth something chronic.  So it is unsurprising that VNC is faster than using a video card.  For 2 reasons:

1. Xvnc doesn't use the graphics hardware.  It renders to an in main memory buffer.  No data is travelling on the AGP or PCI buses. 
2. The VNC protocol is a polling protocol that only refreshes the screen when necessary and only transmits what has changed.  It is also asynchronous to the video updates.  This means that only a fraction of the video updates are being sent over the ethernet wire.

Oliver Jones » Director » » +64 (21) 41 2238
Deeper Design Limited » +64 (7) 377 3328 »