Taiwanese Toast…OR…Why You Can’t Overclock Your Asus P4S8X Very Well

An in-depth look – Brett Wasserman

Many of the Overclockers.com readers have kindly sent me e-mails regarding my recent reviews of the Iwill P4HT and the Albatron PS845PE Pro2. One recurring theme in these messages was that it’s interesting to look at and exhaustively test GOOD motherboards, but it would also be interesting and informative to look at examples that don’t perform so well.

The particular motherboard that readers were curious about more frequently than all the others was the Asus P4S8X. This motherboard employs the Sis 648 chipset and includes many good on-board and overclocking features. While this board comes from Asus, the largest motherboard manufacturer and known generally for very good products, many comments in various message boards seem to indicate that this board is “difficult” to work with.

With the help of the guys from

we borrowed a stock, retail P4S8X for testing. As those of you who’ve read my previous reviews know, I take an approach to testing that does its best to isolate the particular item under test.

Again I’ll be using the Ultra-X RAM Stress Pro card to test CPU, Northbridge and memory functions and I’ll be using the PHD PCI test card to test the remainder of the motherboard. As we’ve seen in the past, these cards function without any operating system installed, exhaustively test each individual system function and clearly identify the specific failure when we overclock a board to its breaking point.

Another piece of equipment I decided to use for this write up is the VICS RD2 PC Geiger. This device does two very useful things for us: It will report the POST codes (as is done on boards like ones from Epox with the dual seven-segment diagnostic LED’s), and it will measure and display the operating frequency of the PCI bus.

Our test rig for this review consists of an Asus P4S8X board rev 1.3, BIOS ver 1003A with –

• Unlocked Pentium 4 1.7 GHz Willamette CPU
• 256M Corsair PC3500 CAS2–one single bank module
• Vantec 470 watt Stealth power supply
• Liebert 2.7kVA uninterruptable power source (UPS) with integrated power conditioner
• Triplex Millennium Silver Ti4200 64MB AGP video card
• Vantec Aeroflow heat sink fan assembly with 5800 rpm TMD fan
• 2 YS Tech 80mm speed adjustable 12 Volt fans
• Floppy drive (for loading test script and saving results)

This is the point in a normal review where I would show you screen shots of the BIOS pages and all of the important overclocking functions. Sadly, I cannot do that. Far into testing but BEFORE I took most of the important pictures for this review, the board decided that it didn’t want me to post its review and internally shorted the 12 volt power circuit to ground. Although the board failed, I WAS ABLE TO DETERMINE SOME REASONS for its reported difficulties.
{mospagebreak}

Brett Wasserman

My first test is always to run the subject motherboard with a complete set of default BIOS settings. In this case the default settings are:

• FSB=100
• PCI=33
• CPU multiplier=17
• Mem ratio=Auto
• Mem settings=by SPD
• Vcore, Vdimm, Vagp=default

With these settings, the board unfortunately refused to POST. That’s odd…I would have thought that setting the BIOS at Fail-Safe defaults should ALWAYS result in a runable system assuming that all of the parts are working. Obviously I’m wrong.

After a little more testing, I realized that the Auto setting for the memory ratio always chooses the highest setting — in this case, FSB * 2 = DDR400. This is a poor choice of defaults since none of the board’s documentation makes any mention of memory support faster than DDR333. It also isn’t anything close to a “Fail Safe Default” setting as it’s called in the BIOS.

Realizing that the default memory ratio setting was the problem, I then backed the setting down to DDR333 and the board POSTed and ran all of the RAM stress tests without any failures.

So far, we’ve learned that the P4S8X won’t do DDR400 at default settings. Now let’s turn down the CPU multiplier and the memory ratio and see if we can find some of the FSB/ratio/PCI clock boundary conditions where the board’s operation begins to fail.

More scrutiny of the various settings available in the BIOS revealed that this motherboard has no specific PCI/AGP clock locking mechanism. What Asus SEEMS to have done is to implement sliding, non-adjustable FSB:PCI dividers.

That is, for a given FSB setting, a divisor is chosen to make the PCI clock as close to 33 Mhz as possible. Using a little (very little) math, I determined that the BIOS screen displays a PCI clock that uses a divisor of 3 for FSB settings of 100-119, 4 for FSB settings of 120-159 and 5 for settings of 160 to the board’s maximum of 166 Mhz. (Another oddity for a Pentium 4 motherboard — most other adjustable motherboards can be set up to 200 or 255 Mhz.)

This use of FSB divisors for the PCI (and AGP) clock setting lead me to do a little research into the PCI bus specification in hopes of trying to determine if clock speeds BELOW 33 Mhz could be used. For example we know that 133FSB / 4 = 33 Mhz, which is perfect, but 120FSB / 4 = 30.0 Mhz, which is quite low.

Although I can’t link you to the official specification document, since you must be a member to view it, the relevant passage is:

“2.2.1. System Pins
*CLK in*

Clock provides timing for all transactions on PCI and is an
input to every PCI device. All other PCI signals, except
RST#, INTA#, INTB#, INTC#, and INTD#, are sampled on
the rising edge of CLK and all other timing parameters are
defined with respect to this edge. PCI operates up to 33.0 MHz
(refer to Chapter 4) or 66.0 MHz (refer to Chapter 7) and, in
general, the minimum frequency is DC (0 Hz)”

This specification means that YES, it is OK to operate the PCI bus slower than 33 Mhz, and all the way down to 0 hz. But it also says that we cannot go above 33.0 Mhz, which this Asus board does with many of the FSB choices.

OK…so now that we are armed with the knowledge that the P4S8X won’t operate at the Auto memory setting due to the unattainable memory speed, and the fact that some FSB choices will result in PCI clock settings that might not work, I set out to determine just which ranges of memory speeds and PCI clocks would work. Remember that “working” means looping and passing all of the RAM stress tests and then looping and passing all of the PCI tests, each for at least 12 hours, with ZERO errors.

My thought at this point was to test various FSB settings with each of the four available memory ratios (FSB*1, *1.33, *1.66 and *2=auto) and create a chart of the resulting PCI clock speed and a notation of whether the tests passed or failed. I was expecting to find that a few FSB frequencies would create PCI clocks above what would function, but these would be a very small set of the choices.

My reasoning was that the application of the sliding memory ratios would mean that there would be three different FSB ranges that worked well (PCI frequencies closer to 33 Mhz) instead of one long linear range where operation failed at one particular FSB/PCI setting.

I was AMAZED when I made the measurements below:

17	100	auto	33.6	3	33	fail	no post
17	100	166x2	31.5	3	33	pass	
17	100	133x2	33.6	3	33	pass	
17	100	100x2	33.6	3	33	pass	
10	100	auto	33.6	3	33	fail	no post
10	100	166x2	31.5	3	33	pass	
10	100	133x2	33.6	3	33	pass	
10	100	100x2	33.6	3	33	pass	
10	133	100x2	33.6	4	33	pass	
10	133	auto	33.6	4	33	fail	no post
10	133	166x2	33.6	4	33	pass	
10	133	133x2	33.6	4	33	pass	
10	137	137x2	34.2	4	34	pass	
10	137	171x2	34.2	4	34	pass	
10	137	102x2	34.2	4	34	pass	
10	137	auto	34.2	4	34	fail	fail in RST Pro
10	150	112x2	37.4	4	38	pass	freq fluctuates +/-0.1mhz
10	150	150x2	37.4	4	38	pass	freq fluctuates +/-0.1mhz
10	150	187x2	37.4	4	38	fail	freq fluctuates +/-0.1mhz/fail RST
10	150	auto	37.4	4	38	fail	fails on C2 in POST
10	160	120x2	40.0	4	32	fail	no post
10	160	auto	40.0	4	32	fail	fails on C2 in POST
10	160	160x2	40.0	4	32	fail	fail in RST
10	160	200x2	40.0	4	32	fail	no post
10	160	240x2	40.0	4	32	fail	fails on C2 in POST
10	163	122x2	40.7	4	33	fail	fail in RST
10	166	100x2	41.5	4	33	fail	fails on C2 in POST

Take a look at the frequencies that I measured (PCI Act) for the PCI clock vs. what the BIOS screen reports. THE BIOS SCREEN IS LYING ABOUT THE PCI FREQUENCY!

The measurements I took indicate that only divisors of 3 and 4 are actually applied. The PCI divisor of 3 is used below 133 Mhz and 4 is used from 133 Mhz to 166 Mhz.

I think that this is the reason for all of the weirdness and questions people have had with this motherboard. The BIOS indicates that FSB settings of 160 and above should provide a 33 Mhz PCI clock, but in fact they measure at ~40Mhz — certainly too high to function reliably, if at all. It’s hard enough to arrive at a good, stable overclocking configuration when you can’t lock the PCI/AGP clock, but it’s impossible when the BIOS is not telling you the truth.
{mospagebreak}

Brett Wasserman

To recap a bit, we’ve learned that:

• The actual PCI clock is FSB/4 for FSB settings above 132 Mhz and FSB/3 from 100 – 132 Mhz
• Auto (*2) memory ratio at any FSB setting doesn’t work

The reason for the former is that the BIOS screen is mismatched with the settings that are being applied to the clock generator chip. This can be corrected with a BIOS update — either the right way by actually adding the /5 divisor, or the wrong way by correcting the reported PCI clock speed. (I hope Asus reads this.)

In an effort to see if Asus had fixed this egregious BIOS error, I found and downloaded their newest version: 1004. The only release note about this version was a cryptic statement saying “Modify the Vcore voltage setting”. Obviously this has nothing to do with the PCI clock frequency, but I downloaded and applied the new BIOS anyway. (Maybe Asus had snuck in a fix without listing it in the release notes.)

No luck. After applying the new BIOS, an FSB setting of 166 still results in a PCI clock of 41.5 Mhz, which indicates a divisor of 4. The BIOS screen still reports a PCI clock of 33 Mhz.

Now here’s where the fun starts… After applying the new BIOS and checking the PCI clock, I switched the PC Giger back to the position where it monitors the POST codes and noticed that it was hung testing the RAM. I pushed the reset button, waited a bit and then heard the speaker emit a continuous series of long beeps. The monitor stayed blank, the CPU fan stopped spinning and I smelled that tell-tale burning electrical smell that we all know and love. This is BAD — the board is on loan from Directron (sorry guys — it wasn’t my fault) and all of my components are in it.

Luckily the only casualty was the motherboard. My CPU, memory and other cards thankfully weren’t damaged. What’s amazing is that the board does start up, POSTs and attempts to boot — all with the 12 volt line totally turned off by the power supply due to a short circuit it’s detecting. Of course, it fails when booting, but amazingly it isn’t totally dead.

As I mentioned above, I wasn’t able to get the P4S8X to work at all using the Auto (or 2X) memory setting regardless of the FSB and CPU multiplier setting. Relaxing the memory timings didn’t help or change the failure symptoms here either. Why doesn’t this aggressive but usually achievable setting work? I think that the answer lies in the picture below:

What are All Those Squiggly Lines?

The lines that you see in the picture are traces on the back of the P4S8x that are between the CPU and the DIMM sockets. I’ve mentioned in previous reviews how important the actual layout of a motherboard (traces, holes, component placement, etc.) is to proper function. This can’t be stressed enough. The best set of components will work poorly, or not at all, if very strict layout guidelines are not strictly followed.

Here’s a quote from Micron’s design guidelines for systems using DDR333 DIMMS:

“Trace Length Matching
±100-mils of memory clock length at the DIMM”

Micron Spec

…and here’s a slightly more detailed quote from E-Insight magazine on the same topic:

“When using unbuffered modules, each DIMM socket requires three clock pairs. The routing of each clock pair includes parallel traces approximately 10 mm apart and at least 25 mm away from any other signal. Standard differential pair routing rules apply. The command and address signals are both synchronized to the clock, and it is important that the clock signals, address signals and the command signals to any one DIMM match in length. The extent of required matching depends on the controller requirements and the timing budget.

Data Lines. The data (DQ), data strobe (DQS) and data mask (DM) signals switch at twice the rate of the clock, and the DQ and DQS lines are bi-directional. Therefore, additional care is necessary when routing these signals. To help maximize timing margins, it is important to keep all signals within each byte lane, matched in length, to within 10 ps to 15 ps.”

Both of these documents are saying the same thing, although Micron’s spec is a little tighter. (100 mils=.100″ and 10ps (distance signal travels in 10 pico-seconds) ~.125″ if the signal propagates at the rule-of-thumb 1ns per foot.) They are saying that the length of the traces on the clock and data lines between the CPU, DIMMS and Sis648 Northbridge need to be of the same length to within ~.1″.

All of those squiggly lines that you see on the motherboard are attempts by the board designers to comply with these guidelines.

They’re trying to add length to traces by NOT using the shortest path between two points. Their attempts are valiant, but every time they make the trace turn, they add two to three pF (pico Farads) of capacitance to it. This added (and mismatched between lines) capacitance changes the timing and transmission characteristics of the trace.

While they might end up with traces that comply with the LENGTH MATCHING specification, they are not matching the overall characteristics of the traces and thus making the board much less reliable at frequencies higher than the specified DDR333.

Why did Asus do this? Was it a new designer who wasn’t experienced enough to create a great design? Was it simply a rush to get the board to market? That’s a little hard to say, but there is one interesting aspect of the P4S8X that might be telling us something.

In various areas of the motherboard, there are solder pads and silk-screened legends for components that aren’t there. This by itself isn’t so unusual. Many manufacturers use the same base circuit board for different versions of the similar boards. This is particularly true of Intel 845 boards of all flavors.

What’s unusual with the P4S8x though, is that there are solder pads and silk screened legends that are unused, but those same parts are placed elsewhere on the board and their legends are incorrect. (There cannot be 2 parts both labeled J10 or C2.)

What this says to me is that Asus took an existing board design (most likely for the P4S533) and very quickly modified its design to accommodate the 648 chipset. As I said above, they might have complied with the letter of the design guidelines, but didn’t end up with a product that runs well above its published specs.
{mospagebreak}

Brett Wasserman

In the real world using locked CPUs, this board’s overclocking capabilities are more geared towards higher memory bandwidth than high front side bus speeds. Realistically, due to the PCI clock divisor issue, you will max out the FSB somewhere around 157 Mhz.

At that FSB, you’ll have a choice of running the memory at DDR235, DDR314, DDR377 and DDR418. I doubt that DDR418 will work since I wasn’t able to get DDR400 (with EXCELLENT Corsair PC3500 CAS2 rated DIMMs) to work, but I suppose with some tweaking of the memory settings, it might be possible.

The features included with the P4S8x are quite extensive and Asus does a generally good job of making them usable right out of the box. All of the necessary cables and headers are included, so there’s no need to search for a compatible SPDIF or Firewire cable. The one thing that’s noteworthy though, is that Asus chose to use a nonstandard connector for the on-board Firewire header and they placed it in a particularly bad location on the board.

If you lose or break this cable, you’ll have a very hard time finding a replacement.

The P4S8X employs a two phase voltage regulator design controlled by the Intersil HIP6302.

This circuit is fully compliant with Intel’s VRM9 specification and will have no problem delivering sufficient power (50A at Vcore) to any current Intel CPU. (I’ll be writing more about the differences between VRM implementations in an upcoming article.)

In my never ending quest for insight into the power usage of PC components, I was able to spend a little time (before toasting the board) seeing just how much power I could get my test components to consume.

DC curent measurements (simultaneous Prime95, disk defrag, CD burn)
Motherboard:

Once again we see that:

• Two phase power is fine for current CPU’s
• Even under stress (and I have to say that simultaneously doing a disk defrag, burning a CDR AND running Prime95 is about as stressful as a system is gonna’ get) we didn’t break the 100 watt mark in total power draw from the power supply. (Maybe next time I’ll add loop 3D benchmarks too.)

• Unless you’re using a Peltier cooling system or using multiple 12 volt pumps in a water cooled system, a 300 watt power supply is sufficient.

Summing up my feelings about this board has been difficult. The problems — BIOS not reporting the correct PCI clock, inability to run at AUTO memory ratio and most of all, the board’s failure — lead me have a rather negative opinion of it. I just have a fundamental problem with a board that won’t operate at defaults right out of the box.

All of that being said though, I do think that one can use the P4S8X as long as they are aware of the issues that I reported in this review and aren’t expecting to reach FSB speeds above the high 150’s. For about the same cost though, I would prefer using an Intel 845PE based board right now.

I’d like to thank the guys at Directron for loaning us this board so that we could try to shed some light on user’s troubles. I’d also like to apologize and ask for their forgiveness for toasting the board. I’m hopeful that Asus will happily replace it. If not, then I guess my payment will be on the way.

Brett Wasserman