AGP, vs PCI, and throw in a little UPA was Re: [SunRescue] NetBSD v. Linux v. Solaris 2.51

[Gregory Leblanc]

On 07 Dec 2000 01:39:18 -0800, Christopher Byrne wrote:
> James Lockwood Sent: Wednesday, December 06, 2000 23:41
> >The Sun PCI implementations also offer multiple PCI busses.  On the Ultra
> >5 and 10 the second 32-bit/33MHz bus is used only for onboard devices
> >(thus avoiding contention with expansion cards) but all other Sun PCI
> >based systems offer at least 2 separate busses for cards (the AXi has 2
> >groups of 3, for example).
> 
> The multiple bus architecture certainly greatly improves resource allocation
> over the standard PC implementations, especially for those systems which
> have 64 bit slots, but it doesn't change the fact that with todays ultra
> high bandwidth devices, PCI is still a big limitation.
> 
> Moving to 64 bit and 66 mhz helps a lot, increasing maximum bandwitdh to
> about 500 MB/s, but that is stil inadequate for the upcoming generation of
> I/O controllers with speeds into the Gigabyte per second (FC-AL is expected
> to be available at that speed within 18 months). What we really need is a
> revolutionary rather than evolutionary jump in peripheral and component
> interfaces. Some people are suggesting that serial I/O is the way of the
> future, but I don't find myself convinced.


What are they calling the new bus that's supposed to be replacing PCI?
InfiniBand or something?  From the datasheets that I've seen, it looks
pretty slick, but I'm not sure how much of what I've heard is still
under NDA.  Probably not much, as there are actual designs based on the
spec already...  InfiniBand certainly looks to be a good replacement for
PCI on the bandwidth side.  My biggest question is how many machines can
saturate a couple of Ultra160 SCSI cards and a couple of GigE cards
simultaneously, and what do those machines look like?  90% of the nodes
on our lan are still running at 10BaseT, so GigE is a ways off for us.

    Greg