[Open-graphics] A few words in favor of multiple chip architecture

James Richard Tyrer tyrerj at acm.org
Tue Aug 22 03:38:58 EDT 2006


Various threads recently make me think that we should consider multiple 
chip architecture for the video board.

My current idea is here:

	http://home.earthlink.net/~tyrerj/files/OG/NB04.pdf

	The considerations are:

	GPU will probably improve over time while some of the other
	stuff will remain constant.

	Analog video is going to become less and less popular to the
	point that cards that don't support it are going to become
	popular.

	There are currently 4 PC bus interfaces: AGP, PCI, PCIe, and
	HTX, as well as PCI-X which doesn't seem to be popular enough to
	support.

	We might want to add a CPU for X on the card.

	We might have other stuff that would need to connect to the I/O
	bus.

	We might want MPEG decoding on the card.

So what I have is basically a custom NorthBridge that should be 
realizable with existing cores.  It should contain:

	Memory controller for DDR & DDR2 memory, 128 MB to 512 MB.

	VGA compatible display controller.

	Video refresh controller.

	Front side bus controller.

	I/O bus controller.

	Bus arbitration and crossbar.

Optionally (these could also be in the "VIDEO" block):

	Hardware scan converter

	YUV - RGB converter

The HW scan converter is only needed for analog video so it might be 
best to have it in a separate chip.

OTOH, the YUV -RGB converter would be needed for a digital video display 
so it would be best in the Bridge chip.

The Bridge chip would have a long product life and might even outlive 
new developments.

Other accelerator stuff can be added to the front side buss including an 
MPEG decoder.

But the really nice idea is here:

http://home.earthlink.net/~tyrerj/files/OG/NB05.pdf

Need more performance, just add more GPUs.

One issue with multiple GPUs is that they need to have different mailbox 
addresses so there would need to be a way to map their control registers 
to different places in I/O space.

And you could use the same GPU chip to build a super computer.

Also, with an interface chip (a different Bridge chip containing a video 
refresh controller and [if needed] a VGA core), the GPU could also be 
used with a single memory embedded system.

-- 
JRT




More information about the Open-graphics mailing list