On 10/7/05, Attila Kinali <[EMAIL PROTECTED]> wrote:
> On Fri, 7 Oct 2005 08:45:14 -0400
> Timothy Miller <[EMAIL PROTECTED]> wrote:
>
>
> > > How about using here 64bit for both address and data ?
> > > Or 64bit data and 48bit address (though i doubt that these 16bit
> > > would make much difference) ?
> >
> > Why do we need that?  Is there something special about PCI Express?
> > Ok, I see your point on the address, but keep in mind that OGA's
> > internal address space really only needs 28 bits, and if we have a
> > PCIe controller that can do 64-bit addressing, it's going to translate
> > down.
>
> Yes, i realized too after i send the mail.
>
> > The only reason for 64-bit data is speed.  If we had a 64-bit
> > PCI interface, it might help, but even then, we could just double the
> > data rate on the local bus.
>
> How much up can the data clock go? I mean a single lane of
> PCI-E IIRC gives 2.5Gbit/s, which is about 80MWords/s at 32bit word size.
> This would mean that we would need 80MHz per PCI-E channel.

Frankly, the first support we have for PCIe isn't going to be stellar.
 Mind you, there's only so fast the GPU can run, so that may not be a
problem.  But basically, we intend to put a PCIe to PCIX bridge on a
board and let it translate.  That'll add some latency, but our
dependency on DMA will mitigate that to a large extent.  That's one of
the reasons for supporting PCI-X.  PCI-X typically runs at 133 MHz (At
32-bit, that's 533 MB/sec).  There's also a PCI-X 266.  If the bridge
chip allows us to clock the PCI-X side arbitrarily, and it supports
speeds faster than 133, we'll overclock it to as fast as either chip
will handle.  I would be surprised if TRV10 could not run its
controller at 266MHz or faster, but we'll see based on what I can get
out of the FPGA version.


>
> > > What about JTAG ?
> >
> > Yeah, that's part of what I was missing above with the control
> > interface.  Oh, and we need to not forget some headers to reprogram
> > the Lattice.
>
> Maybe it would be a good idea to have some pins to be used
> to represent some internal states for debugging.

Sure, and we can use unassigned I/O headers for that.

> Apropos, have you thought about a way to clock the whole card slower
> so that it could be debugged with a cheap digital analyser?

Well, we can clock OUR stuff however we want.  The issue is
underclocking the PCI bus, which would require hacking the
motherboard.

_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to