PCI latency and VGA drivers

[Graeme Gill <graeme@colorstar.com.au>]

For those who are interested in the real time performance of
PCI, here is an interesting tit-bit.

In my experience of testing popular PC based PCI systems,
one of the main culprits in long bus access latencies is
the reliance of video drivers on the PCI retry mechanism
to cope with command/data fifo full states in the PCI
VGA chipsets. (See John R Pierce's reply to Brian Sassone
on 15 Oct 1996, "Re: Problem w/ Matrox Millenium")

This problem continues, even with the latest and greatest
AGP based video cards. For this problem to occur, I suspect
you need to be on a system with a host chipset that
blocks non-host PCI master access to system memory,
when the host CPU is stalled by the VGA access retry
cycles (ie. an Intel host chipset).

The Matrox cards have had a reputation with me for
being one of the worst offenders in regard to
this problem, but they seem to have redeemed themselves
somewhat with their latest Windows drivers, by
providing a mechanism in their performance tuning setup,
for turning off "Use automatic PCI bus retries".

This makes a world of difference to the bus latency. It
drops from milliseconds to microseconds, and should
improve the prospects of real time peripherals (video
capture, audio I/O etc.) actually working.

[Now if only they would make sure this "feature" can]
never be turned on!]

Greg Hanssen's article on this subject
(http://web0.tiac.net/users/auldwrks/vgakills.htm), hints
that Tseng Labs and S3 may also have ways of turning
this mis-feature off, but his information seems to
only be about Windows 95 drivers. Anybody know
anything more about this ?
(Windows NT drivers are of particular interest to me.)

Given that fallout from this problem seems to be felt
fairly widely amongst audio card users, I'm a little
surprised that someone with some weight to throw around,
and a stake in their systems working properly (like
Microsoft or Intel ?) hasn't kicked the graphics driver
writers into line.

Graeme Gill.