PCI Bandwidth: Intel (BX/815) vs. Via (KX/KT133)

[subhuman]

Here's the deal
If you can help me tweak an ASUS P3V4X to give the same, or better results for PCI Bandwidth as my ASUS CUSL2, then I'll send you a box of 3 (three) Unused Intel Retail Fans, and one of these SocketSinkers for FREE.

I will be performing the test, as I do not know of any other way to show the issue than to use the motherboard with my Pulsar soundcard which has zero latency, meaning I can run the devices on it in real time. I promise to be 100% honest with my results, I am serious about finding the root of the problem and spreading the cure, if one exists (I'm very skeptical myself).

Previous Results and Findings
On my ABIT BX6 or ASUS CUSL2 (Intel BX and 815 chipsets, respectively), I am able to get 6 (six) SonicTimeworks Reverbs without PCI Overflow errors. On the CUSL2, I can even copy a bunch of large files without getting errors (due to the IDE ports connecting to the ICH2 instead of using the PCI bus as on the BX chipset).

On the ASUS P3V4x, and even athlon-based VIA KX/KT133 solutions, I am only able to get 2 (two) SonicTimeworks Reverbs before I get PCI Overflow errors.

Hints and links to help you
I am only able to get 5 reverbs on the BX/815 chips until I go into the BIOS and change the PCI Latency timer up to 128 or higher. On the CUSL2, I have it maxed out at 255. This seems to help performance (gains me a reverb or 2).

A reverb is basically a set of 30 or so delay lines (and other items of course, but the PCI intensive item here is the delays), meaning, one sound is sent from the soundcard, across the PCI bus, to main memory to store each of the delays.

Here is a link to the topic that spawned this thread.

Finally, here are a few links from major audio companies showing the problem, since most people on this board have no idea what I'm talking about.


* K7 Athlon and Steinberg (makers of Cubase)
* Creamware, makes of Pulsar/SCOPE/Elektra/Luna cards
* Steinberg's Audio card. Note that they mention "a half-AMD half-VIA chipset" they are refering to the AMD760(ddr) chipset, which uses the VIA Northbridge. I have obtained test results that show this chipset is indeed better than any previous VIA chipset, but still about 10% behind the 815 chipset in PCI Bandwidth tests.

 

[subhuman]

Concurrent PCI and PCI Latency timer don't seem to help the ASUS P3v4X (ie: VIA KX133A chipset).

Any other ideas?

The only thing I had to tweak on the BX/815 chips was setting PCI Latency timer higher, which gained me one reverb.

 

[subhuman]

^bump -- someone must have some ideas, tweaks to try. Everyone seemed to have an opinion before, let's hear it.

 

[Peter]

VIA chipsets have a highly adjustable PCI arbitration scheme, with the one commonly selected giving the CPU and south bridge chips high priority. You need another setting that gives more priority to other PCI masters.

See if your board's BIOS offers any tweaking in that area.

Regards, Peter

 

[subhuman]

Peter,

Could you be more specific? The only BIOS settings I have found which relate to PCI Bandwidth are "PCI Latency Timer" and "Concurrent PCI."

I noticed in the ASUS P3v4X manual it mentions the board (presumably chipset) supports Concurrent PCI, but there isn't any mention of having to enable it, or even the possibility that it can even be disabled at all. It doesn't show up in the manual for the BIOS, nor the BIOS anywhere, just listed as a feature.

On my P3BF/CUSL2 boards, Concurrent PCI is always enabled. Why would you ever disable it?

 

[Peter]

Concurrency features enhance PCI bus availability. No need to turn them off.

As for the PCI bus arbitration tweakability in VIA chipsets, if your board's BIOS doesn't offer any settings like "CPU priority" or "Enhance CPU performance" or whatever, then _you_ can't help it as a user (other than by obtaining a data sheet from VIA and tweaking the register set yourself ). Ask ASUS.

Regards, Peter

 

[subhuman]

Okay!! well, there you have it. VIA boards have basically NO PCI BANDWIDTH. I have tweaked every option in the ASUS p3v4x BIOS, and no matter what, it doesn't perform even 1/2 as good as a BX or 815 chip as far as PCI BANDWIDTH.

The offer is still open though, and I'll give you 4 Intel fans and a SocketSinker if you can help "fix" a motherboard with the VIA chipset -- only for 3 more weeks (or less) cos at that point, we're selling it.

Just tired of dealing with lack of performance...

 

[Peter]

So does the ASUS BIOS offer any settings regarding CPU priority over other PCI bus masters or doesn't it? Any other settings will not improve on your problem.

Don't blame the chipset - all we see here is that ASUS seems to have chosen very high CPU priority. That favors benchmark results, and easily gives a board those .5 percent more "performance" over the competition that the gaming vultures love to settle on, but it totally messes up on PCI bus master throughput. Hell, that's why VIA made that adjustable - only that BIOS engineers would have to make that adjustment available to the user ...

Recently I've seen a report from someone whose CPU priority was high enough to even hurt PCI IDE bus master performance. Luckily, that BIOS had it selectable - that guy turned "High CPU performance" off and instantly, PCI bus master performance was fine.

Regards, Peter

 

[subhuman]

Peter, if you have the time, check out the P3V4X manual (linked above). I've also had the same problem with the KX133 chipset on the Gigabyte board, and also on an MSI KT133/TBird 800mhz combo. I ended up selling both, thinking the ASUS might have more options to tweak.


The manual will who you all the options that I have available. If there is some settings in a program like SoftFSB or similiar, and it requires some Hex values, I will give it a try. Sorry, I am just really frusterated, everyone says there is "no problems with the VIA chipsets" but I'm having many. They are there and real, just most people must not push their PCI bus too hard.

 

[Peter]

You'd have to fiddle with the chipset registers in the Host Bridge part ... "Arbitration Mechanism" and "Master Priority Rotation Control" are the things to look for.

Unfortunately for you, the datasheets for the newer chipsets aren't publicly available - ask VIA if you can have the one you need (694X part). This normally requires signing a non-disclosure agreement but is free of charge.

However, with VIA not changing stuff just for the fun of it, I'd assume that features like these have remained unchanged from earlier chipsets (like the original 691 Apollo Pro) whose full data sheet is available on www.viatech.com. Assuming this is correct, device 0 registers 70h through 76h is your area of experiment. Note, getting things wrong may corrupt your data or OS installation - do this on a backup copy of your working set.

Regards, Peter

 

[subhuman]

P3V4X going for sale this week. No more VIA... Note that I have had the same PCI bandwidth problems with ASUS A7V, Abit KT7, Epox Socket A board, ASUS P3V4X, Gigabyte GA-7ZX. Intel and AMD boards with the VIA chipset, it doesn't matter, brand doesn't seem to matter. No settings I have found in any BIOS makes any difference. (I've sold all these boards but still have one P3V4X left)

If VIA has different arbitration schemes as you mention (and, thank you for any and all your help/advice!), then I would expect that ONE of these boards would have a different scheme? Or, even, a way to change the scheme? But, no, that's not what I have found...

 

[subhuman]

Well, I never did find a solution to this problem. I've been strongly suggesting to anyone needing a computer capable of the MAXIMUM pci bandwidth to stay away from the VIA chipsets. I spent WAY too much time tracking down the problem and trying voodoo solutions to get the same performance my ASUS CUSL2 and ABIT BX6 boards had right out of the box.

I find it frusterating that most people have no clue what I'm talking about, and that PCI Bandwidth isn't measured in any review. In fact, I am not sure of a way to test PCI Bandwidth besides using a very expensive PCI Bandwidth Testing hardware unit ($8000) which I can't find a link to right now (it's a product made to test the PCI bus timings for PCI card mfgs).

Anyway, I did find some interesting information about tweaking the PCI Latency Timer, in fact, IT DOES INCREASE PERFORMANCE like I have found in my own personal tests, check it out:


Other settings

PCI-latency timer

Each PCI slot has a certain number of clock cycles for uninterrupted access to the system bus / CPU. Since each access also involves initial latencies (penalty cycles), the ratio between idle cycles and active cycles is better if the number of bus cycles (PCI-latency) is increased. The main argument for abandoning ISA slots has been that in a system with ISA devices, the PCI latency cannot be increased beyond 64 cycles. Doubling the latency from 64 to 128 cycles allegedly resulted in an overall system performance increase of 15% as reported by Boot Magazine (May 1998). We have never been able to replicate these findings, however we do see a performance increase by stepping up the latency from 32 to 64 cycles.



source: lostcircuits

I have since obtained results using the AMD760 DDR reference platform, and although the AMD760 chipset DOES SHOW SIGNIFICANT PCI BANDWIDTH IMPROVEMENT, it is still not quite on par with the Intel BX or Intel 815 chipsets. Note that the AMD760 chipset DOES use a VIA NORTH BRIDGE, so this would lead me to believe the VIA SOUTH BRIDGE is the culprit here.

Also note that the 815 chipset with the ICH2 hub seems to alleviate PCI bottlenecks a little bit, taking (some of the) USB ports and IDE ports off the PCI bus directly.

If any other geeks have light they can help shed on this subject please help.

Mostly, I would be interested in quantifying my results, and making a way to duplicate them with minimal hardware (not everyone can have an $1000 audio card to test this, obviously). Any ideas to test PCI Bandwidth?

BTW, my P3V4X is forsale in the Forsale forum Great board for the wordprocessor/gamer type.