RE: 64-bit on-board devices on 32-bit PCI bus

[Neal Palmer <neal@dinigroup.com>]

Weng,

  but if the motherboard chipset has only a 32-bit PCI connection, and the 
device ONLY talks to host SDRAM (through the motherboard chipset), then it 
isn't worth the extra wires to make the 2 devices be 64-bits.

-- Neal

On Mon, 5 Aug 2002, Weng Tianxiang wrote:

> Hi Neal,
> I disagree with you.
> 
> PCI Specs 2.2, page 101:
> 64-bit transactions on PCI are dynamically negotiated(once per
> transaction) between the Target and Master, ...
> 
> If a motherboard is stick to your option 1, no ways for other 64-bit
> boards in a system to do any 64-bit transaction.
> 
> A motherboard should use your option 2 and IT WILL IMPROVE SYSTEM
> PERFORMANCE for ANY 64-BIT BOARDS.
> 
> Weng
> 
> -----Original Message-----
> From: Neal Palmer [mailto:neal@dinigroup.com] 
> Sent: Tuesday, July 30, 2002 11:54 PM
> To: Venkateshwarlu V
> Cc: Anand.Kuriakose@smartm.com; pci-sig@znyx.com; Gopal Savanur
> Subject: Re: 64-bit on-board devices on 32-bit PCI bus
> 
> 
> 
> Anand,
> 
>   You are making a motherboard right?  You have 2 options, both of which
> 
> are valid:
> 
> 1) connect ONLY the 32-bit signals between all 3 parts.  Keep the 64-bit
> 
> signals disconnected (no connections between any devices).  Do not 
> assert REQ64# during RST#.  The two 64-bit devices will automatically 
> drive the upper 32-bits of data and CBE to some non-floating state.
> This 
> doesn't require any extra resistors.
> 
> 2) Connect the devices with the full bus width.  And put down some logic
> 
> to drive REQ64# active during RST#.  You will have to add pullup
> resistors 
> for upper AD and CBE lines.  But note that this will NOT improve system 
> performance because the SCSI and network cards will ALWAYS access host 
> memory which is connected to the PCI bus through a 32-bit chip (i.e. all
> 
> of your transactions to host memory will have ACK64# deasserted).
> 
>   So, you should probably just go with option #1 (unless you are going
> to 
> write special SCSI or network drivers to send the data to a different 
> 64-bit card plugged into an expansion slot).
> 
> -- Neal
> 
> On Wed, 31 Jul 2002, Venkateshwarlu V wrote:
> 
> > ** Proprietary **
> > 
> > Hello Mr.Anand,
> > I will try to answer ur questions here.
> > 
> > Should i provide on-board pull-ups on the upper 32-bit extension or 
> > just leave them unconnected? U should not provide any on-board 
> > pull-ups, incase if u use same extension board for 64-bit connector, 
> > there are pull-ups for all the data lines on mother-board also, due to
> 
> > two pull-ups on each data line, there will be more pull-up current. So
> 
> > upper 32-bits must be unconnected.  I personnaly feel that the 32-bit 
> > extension of these individaul devices should be left unconnected 
> > (hoping that the devices will take care of preventing the bus from 
> > floating). The device cannot take care of preventing from floating  
> > The other thing i thought if i provided a common on-board pull-up for 
> > the 32-bit extension of both these devices is that there could be 
> > probable data contention, because at any time and on any of the upper 
> > 32-bit signals, the two devices could be driving opposite levels. Am i
> right in thinking in this direction?
> > Sol'n: Configure Ur device in output mode(upper 32-bits), when u
> insert it in 32-bit slots(according to REQ64# logic), drive zero/one
> onto the upper 32-bit bus.
> > 
> > 
> > 
> > >>> <Anand.Kuriakose@smartm.com> 07/30/02 11:00PM >>>
> > 
> > 
> > Hi,
> > 
> > As per the PCI spec 2.2, sec 3.8.1, the 32-bit PCI extension has to be
> 
> > pulled on the  motherboard, to prevent them from floating. Heres a 
> > situation where the motherboard has a 32bit PCI bus . There are two 
> > on-board 64-bit PCI devices (LAN and SCSI) on the 32-bit bus. Again, 
> > as per the PCI spec 2.2, sec 3.8., we see that these PCI devices 
> > monitors the REQ64# signals during PCI reset de-assertion to determine
> 
> > the width of the PCI bus. In the above case, these devices will see 
> > that the REQ64# signals is de-asserted during reset de-assertion and 
> > hence conclude that the bus width is only 32-bit. Upon determining 
> > that the bus width is 32-bit, these devices will start driving valid 
> > bit patterns on the upper 32-bit extension in order to prevent the bus
> 
> > from floating. So heres the question.
> > 
> > Should i provide on-board pull-ups on the upper 32-bit extension or 
> > just leave them unconnected?  I personnaly feel that the 32-bit 
> > extension of these individaul devices should be left unconnected 
> > (hoping that the devices will take care of preventing the bus from 
> > floating). The other thing i thought if i provided a common on-board 
> > pull-up for the 32-bit extension of both these devices is that there 
> > could be probable data contention, because at any time and on any of 
> > the upper 32-bit signals, the two devices could be driving opposite 
> > levels. Am i right in thinking in this direction?
> > 
> > Thanks in advance.
> > 
> > Anand.
> > 
> > 
> > 
> 
> 

-- 
-- Neal Palmer

The Dini Group
1010 Pearl St #6
La Jolla, CA 92037
(858) 454-3419 x16
(858) 454-1728 (Fax)