Gigabyte GA-686KDX Motherboard

Back in early 2004, I bought this board (including a manual) for an affordable 1€ plus shipping, intrigued by the combination of dual Slot 1 with SIMM slots. It had a short life on store shelves and seems to be officially forgotten by its manufacturer now. Upon testing more than a year later, it turned out to be working flawlessly and was then used as the basis of a (relatively) quiet retro rig (I already had a real noisy one, which is no fun in the long run). My board is a revision 1F, it came with a 1.42 BIOS in 2 MBit.

Some data and caveats

Potential problems indirectly mentioned:

Electronic user manual

You can obtain a copy of the board's user manual in MS Word format from an old mirror of the Gigabyte FTP. Alternatively, I also have a local 7-Zip compressed version of the same file to offer.

BIOS notes and modification

Apparently the GA-686KDX shipped with both 1 MBit and 2 MBit flash chips and corresponding BIOS versions. The latest version for 1 MBit is 1.43, for 2 MBit it's 1.42.
Dating from mid..late 1997, they are new enough to support


Something can, however, be done about some of the limitations with the help of tools like AwardMod to unpack/repack and Modbin (e.g. version 4.50.80) to modify the main BIOS part (original.tmp).
The 1.42 2 MBit BIOS has been modified as follows: The formerly disabled options for memory tweaking, Delayed Transactions and I/O recovery time were made visible and arranged properly, and a GA-686LX4 lent the microcode patches which include the most current versions for all Klamaths (along with some for newer processors). Later I found out that yet newer updates exist for the production steppings C0 and C1 ( is a goldmine), and incorporated these into a trimmed-down cpucode.bin (6-3-2, 6-3-3 and 6-3-4 only) which should be plenty small enough to fit the 1 MBit BIOS as well.
The result has since been flashed with af56.exe and tested successfully. Future changes might include a newer NCR/Symbios SCSI BIOS and a patch for IDE drives of more than 32 GiB (provided you don't mind the performance).

Download the modified 1.42 2 MBit BIOS.

The 1.43 BIOS for 1 MBit was modified in a similar fashion. The trimmed-down cpucode.bin (6-3-2, 6-3-3 and 6-3-4 only) turned out to fit just fine into the more confined space inside a 1 MBit ROM. Additionally, a little blunder has been corrected via Modbin: A chipset register is not set up correctly in the stock 1.43 BIOS, the data at index 0051 should be 0XXXXXXX (as in 1.42) instead of 1XXXXXXX. The bit in question is bit 15 at offset 50-51h in Intel datasheet nomenclature, misleadingly named "WSC Protocol Enable" in the 440FX datasheet while in fact setting the bit disables the WSC protocol (WSC# handshaking is apparently needed in multiprocessor configurations). The 440LX spec update correctly calls this bit "WSC# Handshake Disable".

Unfinished untested modified 1.43 1 MBit BIOS – this is only lacking the memory tweaking options, as I was too lazy to rearrange stuff in the Chipset Features setup. It does feature updated microcode (6-3-x only), corrects the WSC issue just mentioned and enables the Delayed Transaction option along with the I/O recovery time setting.

As for changes from 1.42 to 1.43, "CPU Overheat Monitor" is now "CPU FAN failure monitor" (which I guess still does the same), and "PCI Burst Write Combine" is now shown (this is not shown in 1.42, but enabled by default anyway). Additionally, 1.43 enables operation with the Adaptec AAA-13x SCSI RAID host adapters.

What I'm doing with it

System components

My GA-686KDX (rev. 1F, 2 MBit BIOS, S/N 9741019490) is now installed in a system with the following components:

[Win2k Screenshot]
Screenshot of Win2k running on the system

(The pretty nixie clock can be found on Frank Bemelman's pages.)

Air guide construction

The processors are really hot beasts when loaded and took me a while to tame - I ended up with a custom-made air guide made of the, err, high-tech material called "cardboard" through which the exhaust fan draws out air. The air can get quite hot under full load (up to 45..50C, the 8412L is only rated for ~20 cfm), but since I'm using idle software it's cool when the machine isn't doing much (better for the poor fan, too!). Some pics from air guide construction:

Since air is only drawn in through the open external 3.5" drive bay and the vent holes in front of the internal 3.5" drive bays, the hard drive (spec'd with 10.0 W idle power dissipation) which sits just there does not need any extra cooling - but the case must be kept closed, of course...

Hardware quirks

This board is relatively picky about IDE and floppy cables or rather their fit - ordinary modern keyed ones frequently won't fit, apparently due to a slightly different position of the keying "nose" on the connector. You may have to resort to old unkeyed cables, observing the position of pin 1. Ah, computer stone age...
While we're at oddities, turning on the main (AC) power frequently results in spontaneous booting with the Seasonic PSU; this never happened with the Fortron. A workaround is holding the power button when turning on the power switch. Apparently the power-on sequencing of the Seasonic PSU and the soft power circuitry on the board do not get along that well.

Installation and configuration

The Win2k install went slowly, but smoothly – I selected the MPS Multiprocessor HAL manually, just to be sure. I would not recommend installing the post-SP4 rollup package on such a system, at least I was greeted with an Inaccessible Boot Device error (0x7B) upon reboot. *grmbl* After replacing several files (among them ntfs.sys, win32k.sys, ntoskrnl.exe, ntkrnlpa.exe, mup.sys) with those from the backup, I got the machine running again and promptly uninstalled the rollup thingy.
Update: Apparently it is only ntoskrnl.exe, a.k.a. NTKRNLMP.EXE build 5.0.2195.7045 which is incompatible with this machine, as are later builds including 5.0.2195.7376 (KB979683). I found like one or two reports of the same issue involving multiprocessor rigs on the web (like this one). The 0x7B error appears when ntoskrnl.exe cannot be loaded, but in this case it's the file itself that's at fault rather than the typical inaccessible boot device. The last working build is 5.0.2195.6992 as shipped in update KB885835, while 5.0.2195.7035 (KB890859) onwards do not work. I'd suggest replacing both ntkrnlmp.exe in system32\dllcache and ntoskrnl.exe in system32 – in this order – by ntkrnlmp.exe version 5.0.2195.6992 (wherever you find it, it's dated October 2004) after all the updates are installed. Installing the recovery console onto disk certainly can't hurt in case anything goes wrong. Future reminders of patches KB890859, KB908523, KB920958, KB931784, KB956841, KB971486, KB977165 and KB979683 must be ignored.
Oh, and for some reason the versions of browser.dll in update KB885835 and the post-SP4 rollup have the same build number even though the latter is newer (and bigger). Looks like someone forgot updating the build number. Make sure you have a browser.dll dated 2005.

Later, installing a second NIC got me another such BSOD, which disappeared after removing it again. I then replaced the stock Win2k LSI/Symbios 53C8xx driver with the one to be obtained from LSI Logic, hoping for this to improve things. Meanwhile I figured out that the host adapter seems to be rather picky about interrupt sharing, and that the BIOS assigns the last two PCI slots the same IRQ. With the SCSI HA in slot 4, putting the DMX into slot 5 (the same used for the NIC giving me a bluescreen) resulted in Win2k stalling on the boot screen. Then I swapped SCSI and graphics (the default Win2k driver for the Millennium II does not even insist on using an interrupt at all, apparently Matrox' 2D engines rely solely on MMIO throughout), and all worked well from that point on, with little effect of sound playback on graphics performance.

Looking at the interrupts assigned to the cards by the MPS HAL (as per the MPS table), I'd even say that all 5 slots have different physical INT lines – as already on the GA-586DX (I used to own one a few years ago), IRQs start from 32 in steps of 4, and those used are 36 (SCSI, slot 2), 40 (NIC, slot 3) and 48 (sound, slot 5) – slot 1 (with IRQ 32) is unused, and I figure slot 4 would use IRQ 44 if the graphics card claimed one. Given that, why the BIOS assigned IRQs would still be relevant when the HAL is already loaded is beyond me.

Thanks to the DOS dual-boot, I can also play DOS games now, with passable exhaust temps. A config for some more funky stuff like Desqview/X is now available as well.

System performance

So, how fast does the machine feel? Coming from a dual 500 MHz Coppermine on a P2B-D with a Deskstar T7K250 (on an Ultra100 TX2) and plenty of RAM, definitely usable. Sure things tend to take longer when its gets CPU/memory intensive, but the load times for OOo 1.1.4 and Seamonkey 1.0a still are perfectly acceptable – shows you how important the storage subsystem is. (This used to reside in my main machine not too long ago, btw.)
Only scrolling in Seamonkey wasn't as much fun due to subpar BitBLT performance (slideshow anyone?), something unexpected in a graphics card with the reputation of being very fast, and sadly there is no 3D acceleration at all. Further investigation revealed that the BitBLT and bitmap writing performance could be improved by up to a factor of 3 (!) by using the standard Win2k driver instead of the 5.82 from Matrox. I suspect that the newer drivers do not turn on write combining correctly, at least the BitBLT scores in Dr. Hardware 2003 coincided suspiciously with transfer rates measured in DOS with and without write combining (in 16 bit: 17 MPixel/s with 5.82, 47 MPixel/s with standard; multiply by 2 to obtain data rates in MB/s). Suffice to say that I was much happier with graphics performance afterwards, an upgrade can wait now. As a positive side effect, the previously noticed interference of graphics transfers with reception on the serial ports is entirely gone. It looks like a card using DMA transfers instead of MMIO for data intensive tasks would have avoided the whole issue; maybe trying some PCI Radeon would clear this up.

In DOS, the machine just flies, even without the SCSI BIOS being cacheable. (That only works with ctppro in real mode, and so far I've not found a way to start this app before EMM386/QEMM is loaded.) Can't beat a fast harddrive... To speed up graphics, I'm using Fastvid ver. 1.10 which turns on write combining. This does not go along well with old games using 16 color modes, so I have to disable VGA WC for these. Another problem is that LFB WC does not seem to work once the system is in virtual mode, while VGA WC works just fine.

Sound quality and performance

While I'm able to use hardware accelerated mode with 288 samples per buffer in ASIO4All with the AWE64, the standard ESS Maestro2E driver used by the DMX is only workable in normal mode with 3 kernel buffers; I selected a buffer size of 576 samples (minimum required is about 384). The ASIO output plugin for Winamp (exe version, 0.53) uses 15 buffers. I'm also using a PCI latency of 64 cycles set via Audiotrak's PCI Latency Tool to hopefully kill some remaining crackles. The effort was worth it; at least when using high-impedance (600 ohm) headphones the output quality (with the optional "headphone amp" being off) is very good, rivaling the Aureons in my main machine. Due to the horrible reputation of the card in terms of driver quality and support or rather lack thereof (for example, a number of the card's features, most notably the digital ins and outs, can't be used in Win2k/XP since ESS never delivered a full-blown reference driver for the Canyon3D chip), I could also pick it up cheaply as a welcome bonus. Musical enjoyment is somewhat limited, however, by the racket the rear fan in the plastic fan bracket makes, even after applying some damping mat pieces. Therefore future sound card upgrades appear unlikely.

Fast forward 11 months: I still play with this machine once in a while. Connecting my relatively new Tascam VL-X5 active nearfield monitor speakers, I found that the old DMX isn't a slouch when it comes to the bass department, better than my usual Aureons. However, load during playback turned out to be fairly high even with no DSP plugins, with not mainly Winamp causing a lot of load but in fact the usage of Seamonkey rising dramatically – apparently ASIO4All and this are battling for memory bandwidth. Finding that waveOut playback caused much less load but wasn't crackle proof, I tried the Kernel Streaming plugin written by Steve Monks – and that did the trick. In fact I can still achieve rock-solid playback with a 32K input buffer and a 4x 8K output buffer, and it proved to be nearly impossible to cause stuttering even with two high priority CPUBurn processes running. On top of that, system load is only as high as when using waveOut or rather even lower, which makes sense given that waveOut is wrapped to the WDM interface (i.e. kernel streaming) anyway.

Now there is only one problem left – at 44.1 kHz, things seem to sound a touch cleaner when using the waveOut plugin with SSRC resampler written by Peter Pawlowski (now known as author of Foobar2000) in days gone by, but of course this one is as crackle prone as the regular waveOut plugin and causes quite some load when resampling. Why resample? This card runs its sound processor and A/D and D/A converters at 48 kHz, with a resampling IC (nominally "inexpensive", whatever that means by Burr-Brown standards) being used to convert other sample rates – either the ESS Canyon3D did not provide enough DSP power for the quality that was to be achieved, or they wanted to keep its performance constant whatever the sample rate.
A clean measurement of output quality proved tricky, as recording quality via the DMX turned out to be very bad in terms of distortion. The input of the AWE64 Value isn't of exceptionally high quality either (plus resampling is needed for 48 kHz, which would be carried out by the kernel mixer in this case), but nonetheless a comparison of the results at 44.1 kHz with those obtained by first resampling to 48 kHz via the aforementioned waveOut plugin with SSRC while playing back a prerecorded test file showed obvious differences. The THD plot of the plain DMX -> AWE64 chain (RMAA rec & play @ 44.1 kHz) not only showed noticeably stronger odd-order harmonics but also a number of anharmonics approaching -80 dBFS which are not present in the setup using Winamp with SSRC (44.1 -> 48) for playback and RMAA (44.1 kHz) for recording. And while the latter still shows more IMD than I'd like, there are much less strong and widely distributed anharmonics and general "fuzz" in the high frequencies. This is why I'd always prefer a sound card with native support for the sample rates one considers important (and 44.1 kHz certainly is), at least as long as it can't throw as much computing power at resampling as an X-Fi.

A while back I also tried various newer OOo versions (2.0.3 and 2.2), where I found that loading an OpenDocument spreadsheet from the recent documents list could lead to an infinite loop and freezing, and continued keyboard scrolling could lead to the system failing to keep up and the whole spreadsheet flickering as scrolling continued more slowly (annoying!).
The first issue was found to be not present in a 2.3 dev build, with the previously slower loading times having been accelerated to the point of being about equal to those of the same document in OOo 1.x format, but while loading times have remained fast in 2.4.1, the infinite loop problem is present again. Maybe it's some kind of compiler optimization that does not play well with the antiquated Klamaths.
The scrolling issue has been resolved at the cost of slight display artifacts, although 1.1.x still is faster in this regard.

In the end I just got OOo 1.1.5, which runs perfectly fine and opens OpenDocument files at that – looks like 2.x still needs some tweaks here and there to be usable on such relatively dated machines. (I've had no such issues on other, more modern ones.)

Web browsers

Round one – July 2008

The following modern-day web browsers were tried on this machine:

Safari did not get along too well with its operating environment. During installation already, component registration failed for a DLL, and while displaying my own home page worked fine, eBay or the preferences dialog caused a crash of the application. Perhaps some DLL requires SSE or something. Due to the sluggish scrolling (font smoothing is apparently done internally and cannot be disabled), it wouldn't have been first choice anyway.

The other candidates install and run just fine.

Comparing Opera and Firefox, the latter obviously is more user-friendly to set up. Install, slap on All-In-One Gestures, Adblock Plus and Link Widgets, create a new toolbar to take the latter extension's buttons – there you are, a very usable modern web browser. It's just a pity that the choices for opening external links have been limited and "open in same tab" is no longer available in the UI (changing the respective settings via about:config is required).
Opera, being the geek browser that it is, offers a very freely configurable UI. The new 9.5 default theme isn't an ergonomic masterpiece though (there's plenty of choice among pretty skins, and a native skin that is much better already, but Firefox comes set up for a very usable native skin by default), and ad blocking requires manually selecting items to be blocked or fussing with urlfilter.ini like in the olden days. (Guess I'm spoiled by subscriptions to filter set files as offered by Adblock Plus.) Not too advanced for the browser that once brought us mouse gestures, a minimum font size and made tabbed browsing popular (three things I would not want to be without today).

When it comes to speed, things look different again. Opera, once touted "the fastest browser on earth", has very response scrolling (I would disable smooth scrolling though) and prev/next jumping. Firefox can't keep up with that, though page load times aren't too different among the two. Opera also seems to make better use of memory and obviously adapts its memory usage to how much it can get (so it takes up a little less than FF after a while). Firefox in turn seems to handle interactive things quite speedily, like :hover.

If you look at this page's footer with Opera 9.5 or 9.51, you'll see that it lacks the left and right borders. Seems to be a new bug for 9.5 which I hope they'll get fixed soon.
There is another system-specific issue in 9.51 – the automatic scrolling activated by a middle click does not update the screen contents correctly, repeating a few lines over and over instead. [Screenshot] This does not happen in 8.54 or 9.27. On another system with newer components (processors, graphics card and OS), 9.51 behaves correctly. Looks like an interoperability issue with the Millennium II driver. Since 9.5x in general appears to have a number of smaller bugs and quirks, I'd recommend staying with 9.27 for now.

Comparing Firefox (Gecko 1.9 with new toolkit) and Seamonkey (Gecko 1.8 with old XPFE), the former has much quicker :hover handling, which makes for a much better feel when scrolling through a content link list like I have at the top of this page. Startup time for Firefox also is remarkably low (but a comparison is not really fair, since Firefox is a browser only while Seamonkey is an application suite that has to load a good bit more XUL on startup). In return, page-wise scrolling (PgUp/PgDn) is somewhat less quick (about 2/3 speed), and sometimes the browser seems to pause in order to "think" for a while, carrying out the user input afterwards. Seamonkey also appears surprisingly modest in memory consumption, though this might even out with some use.

Seamonkey is in a difficult position these days, since version 2.0 which would catch up with Firefox internally won't come any time soon, which is related to a large number of changes under the hood (toolkit, code version tracking system) to be tackled with rather limited manpower. Meanwhile, the 1.1 line without as much public appeal (still a capable browser on the level of FF2, but with a seriously antiquated extension management and an interface basically unchanged since the days of the Mozilla Application Suite) continues. It takes Adblock Plus just the same, and when installing SessionSaver, Mouse Gestures 1.5.2 (the last pre-rewrite version) and possibly Mnenhy (primarily targeted at Mailnews but also ships with a chrome manager that allows removing extensions) and Preferences Toolbar, it's perfectly workable. I would also recommend getting another skin (like Orbit 3+1 with "small icons" subskin). As a camera geek, you might also like FxIF (originally written for Firefox, but works in Seamonkey as well).

Whatever the browser, I would strongly recommend not to install a Flash plugin on a system like this. It's well-suited for slowing down the system and increasing browser memory usage considerably. Sure that means no Youtube videos... but those would be a slideshow anyway. I've been surfing without Flash support in my main browser for years, and there have been very few occasions that would have required pulling out a secondary browser with it installed.

Round two – January 2009

For round two, the list of browsers is:

The last one actually was a non-starter, as it requires XP2 or higher and the installer will therefore complain unter Win2k. (Safari based on the same Webkit engine showed why this makes sense in round one.) When it does work, it's just a slow scroller as Safari though, so wouldn't be much fun anyway.

Seamonkey is doing no worse than in the last test but is slowly starting to show its age. Gecko 1.8 does not support CSS3 media queries yet which I recently put to (hopefully good) use, so real-life page scaling will be a little worse than for browsers that do. Reloading my "Superhets for Dummies" page takes pretty much exactly 5 seconds (with Adblock Plus disabled for the sake of fairness, which seemed to have only a small effect in this case though), and the quick links popup menu (entirely HTML+CSS, btw) reacts quite speedily. Auto scrolling is a little sluggish downwards but fluid in the other direction. (This equally appears when scrolling with the cursor keys , . Apparently this is due to the quick links element which has position: fixed. The page you're reading does not have it yet and scrolls fluidly in every browser. Based on my main machine which is about twice as fast but still hardly a speed demon by today's standards, I assumed this to have no performance penalty, but apparently there is one.) Complete keyboard based scrolling of "Superhets for Dummies" (via PgDn ↓) takes 18.5 seconds, the same back up is equally fast. (Taking out the quick navigation, this becomes 18.0 seconds.)

Firefox 3.1b2, based on a beta version of the latest Gecko 1.9.1 engine, does support CSS3 media queries. Reloading my "Superhets for Dummies" page takes about 6 seconds (5.6 to be precise), a little less if not scrolled all the way to the bottom. The quick navigation reacts a little more slowly than in Gecko 1.8-powered Seamonkey. Auto scrolling was automatically disabled upon first startup, and reenabling it showed why: It's a little sluggish in both directions, so probably some work is required to get it back up to speed. (A page without position: fixed elements scrolls smoothly though.) Complete keyboard based scrolling of "Superhets for Dummies" takes 24.8 seconds, thus deviating noticeably from the generally similar values obtained by the other browsers. (Without the quick navigation, the time drops to 20.9 seconds, still above average.)

Opera 9.63 turns out to be very fast at page loading: A bare 3 seconds for a "Superhets for Dummies" reload – 9.27 still was pretty much in line with Seamonkey at about 5 seconds. (CSS3 media queries have been supported for ages anyway, they pretty much were invented by Opera folks.) The quick navigation feels somewhat more laggy than in Seamonkey, similar to Firefox. (In the old 9.27, the menu extends over the whole available width, and hover handling is very sluggish.) Complete keyboard based scrolling of "Superhets for Dummies" takes 17.3/17.7 seconds down and 16.9 seconds back up (16.2 seconds down for 9.27).
The auto scrolling problems encountered in 9.51 are still present in 9.63. Opera 10 alpha fixes this, it seems to have the somewhat jumpy (at least when position: fixed elements are present) pre-9.5 autoscroll again. One oddity: When you've scrolled to the top of the page, it shows a strange black bar maybe 1em high for a few seconds which then disappears. (Judging by a cross-check on my main comp, this ought to be semi-transparent but comes out pitch black in this case.) I guess they'll get that sorted out until the final comes out. Complete keyboard based scrolling of "Superhets for Dummies" takes 15.2/15.7 seconds, back up you even need only 14.6 seconds. (Scrolling down without quick navigation: 13.8 seconds.)

Gecko-based browsers still allow for some speed tweaking. For example, the time after which the page is first displayed is fixed, but obviously a slower machine will have parsed and rendered less, thus potentially having to reflow more frequently later. Prefs making are noticeable difference are "content.notify.interval" and "content.interrupt.parsing", for example. Here's a little list of what's been used to tweak Seamonkey and later Firefox:

Pref Type Val
content.interrupt.parsing bool false
content.notify.interval int 1000000
content.switch.threshold int 1000000
content.max.tokenizing.time int 3000000
content.maxtextrun int 4095

In a real-life "go to eBay and search for a specific product" test, this noticeably helped the previously stock Firefox in closing the gap to Seamonkey. The latter with an apparently well-tuned Gecko 1.8 still felt a little more speedy though. (Note that none of my tests can claim to be particularly Javascript-heavy.)

Warning: Setting "content.interrupt.parsing" to "false" may break extension installation from the web in Seamonkey 1.1, causing it to stall during download. Things still work for local XPI files. That one took me a while to find!

Winner? If I had to use this machine for web browsing regularly, I'd probably settle for Opera – 9.6x for now and 10.0 when it comes out – and look out for a decent pre-compiled urlfilter.ini. (Besides, a geek browser like Opera is no bad match for this equally unconventional system.) However, a tweaked Gecko (1.8 in particular) is no slouch either.

It be noted that on this system I've had zero luck with the FxIF extension in Seamonkey 1.1.x, while it works fine in FF 3.x and also cooperates with Seamonkey on other machines.

Some benchmark data

These are with tweaked settings and all. DOS benchmarks taken without any memory managers, but with write combining turned on and SCSI BIOS cacheable (via ctppro) if applicable. FSB is usually 66.86 MHz, for a CPU clock of 233.99 MHz - or so WCPUID says.


Cache size/Memory speed info tool 2.65MMX - (c) 1999-2001, LRMS - DJGPP compiled
CPUID support detected... 'GenuineIntel' with FPU TSC MMX
Family=6 Model=3 Step=4 Type=0 Chipset (Vendor/Device ID(Rev)): Intel/1237(02)
CPU clock: 234.0 MHz
Using 32MB physical memory block (alignment = 32)
Bandwidth - MMX linear access test... Read/Write/Copy (MB/s)
Block of     1KB: 1543.0 / 1165.2 / 2020.5
Block of     2KB: 1643.2 / 1202.1 / 2107.1
Block of     4KB: 1682.6 / 1221.4 / 2167.1
Block of     8KB: 1709.3 / 1231.4 / 2133.9
Block of    16KB: 1685.0 / 1236.2 / 336.8
Block of    32KB: 781.8 / 210.0 / 338.4
Block of    64KB: 783.0 / 210.0 / 324.6
Block of   128KB: 783.7 / 209.9 / 317.5
Block of   256KB: 783.9 / 210.0 / 286.8
Block of   512KB: 779.2 / 209.9 / 126.9
Block of  1024KB: 248.8 /  96.6 / 106.5
Block of  2048KB: 248.8 /  79.4 / 100.4
Block of  4096KB: 248.8 /  77.4 /  98.9
Block of  8192KB: 248.8 /  77.1 /  98.2
Block of 16384KB: 248.8 /  76.9 /  97.9
Block of 32768KB: 248.8 /  76.8
Latency - Memory walk tests... ("pointer chasing")
Null size:    3 cycles   1 cycles (overhead 98 cycles)
           steps:   4   8  16  32  64 128 256 512  1k  2k  4k (bytes)
Block of     1KB:   3   3   3   3   3   3   3   3   -   -   - cycles
Block of     2KB:   3   3   3   3   3   3   3   3   3   -   - cycles
Block of     4KB:   3   3   3   3   3   3   3   3   3   3   - cycles
Block of     8KB:   3   3   3   3   3   3   3   3   3   3   3 cycles
Block of    16KB:   3   3   3   4   3   3   3   3   3   3   3 cycles
Block of    32KB:   5   8  13  24  24  24  24  24  24  24  24 cycles
Block of    64KB:   5   8  13  24  24  24  24  24  24  24  24 cycles
Block of   128KB:   5   8  13  24  24  24  24  24  24  24  24 cycles
Block of   256KB:   5   8  13  24  24  24  24  24  24  24  24 cycles
Block of   512KB:   5   8  13  24  24  24  24  24  24  25  26 cycles
Block of  1024KB:  10  20  35  56  56  56  56  57  57  58  60 cycles
Block of  2048KB:  10  20  35  56  56  56  56  57  57  58  60 cycles
Block of  4096KB:  10  20  35  56  56  56  56  57  57  58  60 cycles
Block of  8192KB:  10  20  35  56  56  56  56  57  58  58  60 cycles
Block of 16384KB:  10  20  35  56  56  56  56  57  57  58  60 cycles
Block of 32768KB:  10  20  35  56  56  56  56  57  57  58  60 cycles
This system appears to have 2 cache levels (enabled).
L1 cache (16KB) speed (MB/s): Read=1709.3, Write=1231.4
L2 cache (512KB) speed (MB/s): Read=783.9, Write=210.0
Main memory speed (MB/s): Read=248.8, Write=76.9


(a) with only 3 rows populated, 999 passes (old)

STREAM for DOS v2 by Dennis Lee
1 MB = 1000000 Bytes in the following measurements.

For accurate results, this benchmark should be executed
in a true DOS session, and not a DOS shell under another OS.

   Time        Operation      Mem Speed       Error
   ----        ---------      ---------       -----
 49.66 sec      COPY32       128.75 MB/s       0.1%
 58.88 sec      COPY64       108.59 MB/s       0.1%
 63.05 sec       SCALE       101.41 MB/s       0.1%
 80.85 sec         ADD       118.62 MB/s       0.1%
 76.90 sec       TRIAD       124.71 MB/s       0.1%

These results are comparable with those on the STREAM website.
See <> for info on STREAM.

Type 'streamd ?' for help

(b) with 6 rows populated (now)

   Time        Operation      Mem Speed       Error
   ----        ---------      ---------       -----
 54.48 sec      COPY32       117.36 MB/s       0.1%
 64.76 sec      COPY64        98.73 MB/s       0.1%
 64.32 sec       SCALE        99.40 MB/s       0.1%
 87.71 sec         ADD       109.34 MB/s       0.1%
 85.69 sec       TRIAD       111.92 MB/s       0.1%

Probably the chipset slowed down the timing to accomodate the higher module count. Still, not too shabby; this seems about 10% faster than on an untweaked system with a similar memory config.

System Speed Test v4.78


System Speed Test  Ver 4.78      Report file - created on 03-10-2005 04:57:45

Comments to report: ST318406LW/2941U2W, ctppro etc. 

CPU is in V86 mode: No
Processor         : Intel Pentium(r) II  233 MHz
Original CPU Clock: 233 MHz (66 MHz x 3.5)
Frequency via TSC : 233.86 MHz
External clock    : 66.82 MHz x 3.5
CPUID (TFMS)      : 0634  Codename: Klamath (0.35um)
Microcode Revision: 35
Feature           : 0080FBFFh  MMX(tm):Yes, IA SSE:No, IA SSE2:No
L1 Cache size     : 32 KB
L2 Cache size     : 512 KB
CPU speed index   : 267.99

Total memory size : 192 MB
Memory Bandwidth  : 250.22 MB/s

PCI Video         : Matrox MGA-2164W Millennium II GUI+3D Accelerator

VESA OEM String   : Matrox Graphics Inc.
VESA video memory : 4096 KB (64758 KB/s)

Hard drive 0      : 2231C 255H 63S  17.09 GB
                  Tested in FULL mode
                  Random access time     : 8.87 ms
                  Linear read speed      : 48777 KB/s
                  Min/Max read speed     : 39006 KB/s / 52646 KB/s
                  Buffered read speed    : 64807 KB/s
                  Hard Drive speed index : 1906.53

Mainboard chipset : Intel 82440FX (Natoma)
BIOS vendor       : Award Modular BIOS v4.51PG (06/27/97)
OS version        : MS-DOS 7.10

                ATA/ATAPI Device Information


IDE1---Master-<ATAPI>- FX4830T
          Firmware Revision             : R02E
          Maximum Transfer Mode         : PIO 4, DMA 2, UDMA 2 (UltraDMA/33)
          Selected DMA Transfer Mode    : DMA 2
          Maximum Transfer Speed        : 8467 KB/s (48X)
          Selected Transfer Speed       : 8467 KB/s (48X)
          Cache Buffer Size             : 128 KB 



                Cache/Memory Benchmark
                     Read          Write         Move         Average   
 Cache Level 1    224.41 MB/s  1276.06 MB/s  1137.03 MB/s   879.17 MB/s 
 Cache Level 2    223.79 MB/s   211.08 MB/s   195.63 MB/s   210.17 MB/s 
 Memory           101.36 MB/s    84.55 MB/s    77.56 MB/s    87.82 MB/s 

                MMX(tm) Cache/Memory Benchmark
                   MMX Read      MMX Write     MMX Move       Average   
 Cache Level 1    668.69 MB/s   668.69 MB/s   359.88 MB/s   565.75 MB/s 
 Cache Level 2    527.06 MB/s   211.12 MB/s   206.47 MB/s   314.88 MB/s 
 Memory           250.19 MB/s    77.10 MB/s    76.72 MB/s   134.67 MB/s 

                Memory Modules Information

           SMBus/i2c Host Controller     : Not Present

                PCI/AGP Device Information  [PCI 2.10]

    +----- VEN_8086, DEV_1237, REV_02, SUBSYS_00000000
           Intel 82441FX 440FX (Natoma) System Controller (SU053)
           Class Code : 060000 - PCI to Host Bridge
    +----- VEN_8086, DEV_7000, REV_01, SUBSYS_00000000
           Intel 82371SB PIIX3 PCI to ISA Bridge
           Class Code : 060100 - PCI to ISA Bridge
    +----- VEN_8086, DEV_7010, REV_00, SUBSYS_00000000
           Intel 82371SB PIIX3 Dual EIDE Controller
           Class Code : 010180 - IDE Controller
           I/O Base Range 4              : F000 - F00F (16)
    +----- VEN_102B, DEV_051B, REV_00, SUBSYS_1100102B
           Matrox MGA-2164W Millennium II GUI+3D Accelerator
           Class Code : 030000 - VGA Compatible Adapter
           Memory Base Range 0           : E0000000 - E0FFFFFF (16MB)
           Memory Base Range 1           : E1800000 - E1803FFF (16KB)
           Memory Base Range 2           : E1000000 - E17FFFFF (8MB)
    +----- VEN_1000, DEV_000C, REV_01, SUBSYS_00000000
           LSI Logic LSI53C895 PCI to Ultra2 SCSI I/O Processor with LVD Link
           Class Code : 010000 - SCSI Controller
           I/O Base Range 0              : 9100 - 91FF (256)
           Memory Base Range 1           : E1804000 - E18040FF (256)
           Memory Base Range 2           : E1805000 - E1805FFF (4KB)

                ISA PnP Device Information

   Card  Vendor ID  Serial No  ANSI ID String

      1    CTL00E4   15C12AE4  Creative SB AWE64  PnP

End of report.

The memory bandwidth seems to indicate a 3-2-2-2-2-2-2-2 burst, not too shabby for ye olde EDO DRAM. If Fast Leadoff is disabled and the MA Wait State is enabled (you normally cannot access these settings in an untweaked BIOS), you get ~222 MB/s, probably corresponding to a 5-2-2-2-2-2-2-2 read burst.

SiSoft Sandra MAX3! Standard

Memory bandwidth: 112 MB/s ALU, 113 MB/s FPU. (For reference: Fully tweaked Asus P2B-D at 100.28 MHz with 1 GiB in PC133 modules, 2-2-2 memory timing and an "MA Wait State" setting of "Fast": 385 MB/s ALU, 401 MB/s FPU.)

Lavalys Everest 2.20.405 Home Edition

Read throughput: 255 MB/s, write throughput: 84 MB/s, memory latency: 253.5 ns. (For reference, a PIII-800EB equipped P3B-F at 132.99 MHz FSB running PC133 memory with 2-2-2 timing and a "MA Wait State" setting of "Normal" reaches 670 MB/s, 230 MB/s and 99.1 ns, respectively. A P2B-D with PIII-500Es at 100.28 MHz shows 507 MB/s, 172 MB/s and 124.4 ns. Do note that Asus BX boards seem to be capped in terms of buffered read bandwidth – that does not reflect in real-world performance but the scores of 980 MB/s for an i815E board at 133 MHz and 710 MB/s for an i815EP board at 100 MHz seem more realistic if you intend to compare things.)

Want to drop me a line?

Created: 2005-09-25
Last modified: 2010-04-20