PCI latency mod?

Only for programmers and BIOS gurus with technical questions.
Post Reply
trodas
BIOS Newbie
Posts: 22
Joined: Sun Oct 15, 2006 6:47 pm

From my overclocking experiments I learn that the CPU power in SuperPI and such benchmarks (eg. folding power :twisted: ) goes up when PCI latency go down. Usual settings are 32 - 248 or so, but I want as low, as possible. Better boards can do 16 - but I would more like 8 or even lower - so the question is.
1) how low one can go on KT600 and 400 chipsets, nF2 chipsets
2) how hard is to mod the bios to allow lower values to choose or hard-mod a very low value there

Another related sub-question is - what other latency values are in bioses and if they are twakable :twisted:

All for sake of CPU power, of course :twisted:
cp
BIOS Guru
Posts: 1914
Joined: Mon Oct 21, 2002 9:07 pm
Location: Germany

use WPCREdit
the latency timer is at offset 0D on every pci device.
have a good time crashing your system.
Denniss
BIOS Guru
Posts: 3153
Joined: Thu Mar 21, 2002 8:16 pm
Location: Near Hannover (CEBIT) Germany
Contact:

Why should PCI Latency affect benchmarks ? Modern system do not use PCI that much to have an affect on benchmarks or folding progs, CPU power and memory bandwith is required and that's not affected by PCI latency.
trodas
BIOS Newbie
Posts: 22
Joined: Sun Oct 15, 2006 6:47 pm

cp - I know this proggy. A bit hardcore, but good. However this is a folding machine we talking about. I need it to run w/o user intervention and I clearly did not want log in each day after I power it up to change this settings - not to mention that when I log in once by terminal, I have to log off to power off the machine as well. Power button in that case DID NOT WORK.
From above stated reasons it is clear that I need permament solution - eg. bios mod :wink:
Original DFI bios allow changing this, moded by Hellfire one swaped that for AGP latency. Sure somewhere in the bios is the PCI latency as well, as AGP controler latency (these are tweakable by nForce2 Tweaker - sadly it is buggy and did not start automaticaly as it should on remote machine with XP...)

Denniss - because more free CPU time is left for the bench :wink: And exactly for these reason you mention I need tweak the PCI latency down. It is setup for case where there is many cards in PCI slots. I have none, so, not need for me long PCI latency :wink:
cp
BIOS Guru
Posts: 1914
Joined: Mon Oct 21, 2002 9:07 pm
Location: Germany

that's why there is WCPRSET
Denniss
BIOS Guru
Posts: 3153
Joined: Thu Mar 21, 2002 8:16 pm
Location: Near Hannover (CEBIT) Germany
Contact:

PCI Latency does not affect CPU time, it is an indicator how long a PCI device is able to have exclusive rights on the PCI bus. After the selected time the next PCI device gets exclusive rights (if needed).
cp
BIOS Guru
Posts: 1914
Joined: Mon Oct 21, 2002 9:07 pm
Location: Germany

Denniss: i totally agree with you. and after all that's what the technical documentation says, too. there's no impact on anything but the pci bus and the devices on that bus. i just wanted to supply the tools to change the pci latency. whether it does influence the (measured) cpu speed or not.
trodas
BIOS Newbie
Posts: 22
Joined: Sun Oct 15, 2006 6:47 pm

Of course I cannot agree more with Deniss PCI latency definition:
PCI Latency does not affect CPU time, it is an indicator how long a PCI device is able to have exclusive rights on the PCI bus. After the selected time the next PCI device gets exclusive rights (if needed).
But he is deadly wrong on the results. First at all, there is a reasonably BIG increase of score in even CPU benchmarks like SuperPI or Hexus PI fast and since I did the test myself, I know that this is not any hype around, but true reality.
In fact, it is logical. Look, the exclusive rights also mean stoping other devices using the bus and quess what - that does (obviously) add the time need for complete of some disk-related-requests that are obviously being made during the benches (start Filemon, if you in dubt). Or think, if you can - these requests are mostly loading the check-data to see if machine is computing the bench right and do not produce bogus results (as overclocked machines beyond their limits usualy do)...
Post Reply