| PCI latency......a whole 'nuther can of worms [message #64125] |
Thu, 02 February 2006 23:27  |
![Deej [1] is currently offline](/theme/PARIS_Forum_Template/images/offline.png "Deej [1] is currently offline")
Deej [1]
Messages: 2149
Registered: January 2006
|
Senior Member |
|
|
I'm just having such a gret time in *nativeville* these days. So much to
learn. I've been getting crackling inmy audio when streaming tracks from
Cubase to Paris when using large numbers of UAD-1 plugins. Well, come to
find out, there's more fun to be had in tweakville......
Here's some info on PCI latency.
http://www.soundonsound.com/sos/Oct04/articles/pcnotes.htm
http://www.uaudio.com/webzine/2005/june/index5.html
http://mark-knutson.com/t3/
http://downloads.guru3d.com/download.php?det=951
From what I can determine from reading a few threads about this, PCI latency
is the amount of "wait" time PCI is allocated to communicate with any given
peripheral. A high PCI Latency setting takes more PCI bus time than another
device with a lower setting. Normally, the PCI Latency Timer is set to 32
cycles. This means the active PCI device has to complete its transactions
within 32 clock cycles or hand it over to the next PCI device. As you can
see, a device, like a video card which has a setting of 248 essentially
"hogs" the PCI Bus.
PCI latency timers are a mechanism for PCI bus-mastering devices to share
the PCI bus fairly. A device such as the RME 9652 gains bus ownership and
the clock counts down based on the latency setting. In our case the RME
9652 specifies a clock count of 255 (unlike most devices which accept the
default count set equally for other devices on the the PCI bus). You might
want to check the default PCI latency for the MADI.
In most cases the healthy level for setting this is around 32-64, but can
sometimes be higher for various sound cards or video cards reaching to the
upward amounts of 128.
The 255 requested by RME HDSP cards seems to be wayyyyyy on the high side.
(which is in fact the maximum value available). I understand that setting
this value too low can can interrupt transfers unnecessarily and hurt the
9652's performance, but setting the value too high can cause other devices
to wait longer than they should have too, therefore overflowing their
buffers. this can be really problematic with some network cards (and I'm
using a Marvel onboard LAN so I'll have to check this further.
I used the Doubledawg PCI latency utility to tweak my PCI latency settings
a bit. I dropped the latency of the Matrox G450 to 32, switched my UAD-1
cards from 128 to 64 in the UAD-1 control panel and backed the latency of my
HDSP 9652 cards from 255 to 248. This, plus changing to buffers in my RME
control panel to 2048 has solved my problem with *crackling* of the audio
with 15 UAD-1 plugins using 40% of the available CPU resources of the cards.
Any more than this and the crackling returns so this is definitely an issue
with the UAD-1 cards. Since my UAD and RME cards are in a Magma, tweaking
the settings for the PCI-PCI bridge (perhaps increasing them in this case
while lowering the latency timing on the cards themselves) might be a fix.
It's going to be interesting......and as I suspected.......it never ends.
;o)
|
|
|
|
The PARIS Forums
Search
Help
Members
Register
Login
Home
