I don't think i can take a screenshot of Speedsys or cachechk with L2 enabled. Things are very fussy.
Can try tomorrow again, but good chance it does not work.

Doom at 2x45 with BL3 = 22.8 fps, not bad but also not great.
Certainly below the 26.2 your Chaintech 340SCD accomplishes with L2 enabled.

Oh, that is a pretty big difference in speed. Looks like the L2 is auto disabling? If there is a BIOS setting for L2 speeds, try slowing it down.

When you have a chance, could you show your cachechk/speedsys at 2x45 and 2x50 (if completes) w/L2 enabled in BIOS?

No, no - it is not auto-disabling. I am disabling it. : )
I can try to provide Speedsys with L2 at 45MHz, but for 50MHz will be impossible.
At 2x45 and no L2 cache Doom and Wolf3D are not great, but even without L2 cache Quake 1 is screaming.

pshipkov wrote on 2023-03-06, 00:08:

..... @CoffeeOne […]
Show full quote

.....
@CoffeeOne

.......
As for the "limited" local storage.
Can you fill even 4Gb with 486 class software ?

I installed an EEPROM with XTIDE on the ISA 3com.
Now I have the full 16GB, yay.
So re-installing Windows 98 again (not the first time 😁).
Just thinking, maybe I should get a 128GB CF card? What is the maximum? Windows 98 can handle 120GB as far as I know.

I believe the fat32 limitation is 137 gigs.

Cool about XTIDE.

FAT32 limits, by memory:
file - up to 4Gb
partition - up to 2Gb
volume - up to 32Gb

max volume size for FAT32 is 2TB

Corrected my post.

If you decide to use Windows 95, I believe the max this OS will support, according to Microsoft, is 32 GB. https://www.computerhope.com/issues/ch000276.htm

I looked into this some years back and recall this 32 GB limitation is related to some special use cases with Scandisk. I have always ignored this limitation and used Scandisk in DOS, which was the Internet's recommendation. In practice, I've found W95c to handle 120 GB partitions without issue. Besides the minor scandisk.exe bug, W95 should handle up to around 127 GB partitions, if I recall correctly.

I think the 2GB partition size limit is best versions of Fat16... which maybe used on 95 due to sector size smaller and less wasteful on smaller HDD sizes... or do I mean cluster size... anyway could be several megabytes difference per GB with a lot of small files. One extra game/app install.

So umm, is there a software way to disable FPU? I apparently can't do it in Setup... and an external setup will disable it, but next boot board goes to setup and wont boot until it's saved with FPU enabled. No obvious jumper onboard. Don't wanna unplug it for a few noFPU tests. Does it only really matter on speedsys anyway?

If you don't have a jumper you cannot hard-disable FPU on these old motherboards from BIOS.
Even the BIOS masks it as "absent", the circuitry is there and if the FPU makes the system unstable for example - that will still be a limiting factor.
So, either a jumper, or removal of FPU is required.

---

Something unexpected and interesting.

Two identical 386 hotrod motherboards.
Same CPU, RAM, L2 cache, VGA, IDE used everywhere.

First board overclocks to 45MHz FSB (90MHz crystal oscillator), second one can barely hit 40MHz (80MHz crystal).
Assumed production line fluctuations in quality at play, or just uneven aging (despite both boards being in great visual condition).

At some point the EEPROM of the does-not-overclock-well motherboard died.
Replaced it with the EEPROM from the other board.
Tried the 90MHz oscillator - instant win.

First EEPROM is INTEL 72512J-20.
Second one is TOSHIBA TMM27512AD-20.

Light bulb.

Reprogrammed the INTEL 27512J-20 guy.
Inserted back in - board cannot go past 40MHz FSB again.

Programmed and used ST M27C512.
Boards hit 50MHz FSB (100MHz crystal) no problem.

Expanded the test with two other and different 386 hotrod boards.
Plus 4 more EEPROM chips from different brands - AMD, TI, WINBOND, SGS-THOMPSON.
All of them 27C512 models.

They are not all equal in terms of overclocking. ST and AMD EEPROMs worked best.
That of course is very inconclusive, but until i accumulate enough passion for proper exploration, i will be sticking to ST and AMD chips.

Learning ?
EEPROMs matter a lot for retro overclocking.

I am a bit stupefied that it never occurred to me before.
Wonder how much inaccuracy this will add to the captured metrics from the examined so far 286/386 motherboards ...

haha, I guess you will need to re-do EVERYTHING!

Is someone able to explain what is happening here? I was going to say that it was the EEPROM's read time, but in your example above, both had the same read/response times. As the BIOS is normally copied to RAM (shadowed), I wouldn't imagine that the EEPROM brand would ever matter. In your above tests, were you not shadowing the BIOS?

Most EEPROMs on 386 mobos are predominantly ST brand, but there are others as well.
Going to comb tomorrow through the 386 inventory and will see.
286 boards are too slow for this to be a factor, but i still wish i stayed ignorant.

As for BIOS shadowing - thats for sure, but apparently if the EEPROM itself is not good enough to handle given frequency = no lights to start with.
The Intel chip from previous post causes motherboard not to light up every time with 80mhz crystal - it is on the edge.

This is quite interesting indeed and will require study to narrow down the weak EEPROMs. If you start with the Intel EEPROMs, I suggest obtaining different speed ratings to see if they fail even at their fastest speed. I haven't researched all the speeds, but there should be at least 75 ns - 200 ns. Were there faster, like 45 ns?

Luckily, I normally use SST and WINBOND and have a tendency to remove the Intel pieces.

BitWrangler wrote on 2023-03-31, 15:04:

TL;DR silicon lottery is like a lottery with silicon.

Yah all RAM, ROM, CPU are graded like eggs, small, medium, large, extra large, but each is a range. So you can get a large medium that looks real close to a small large etc... it just didn't quite fit through the hole. When EEPROM graded 200ns 180ns 150ns, that means 200ns parts range from 199.99 to 180.1 ish, ~10% range. Enough to make go/no go differences over a few mhz. So, you may need large samples to tell whether you've got one brand better than the other. Then also it's going to vary over time. There is a target spec, but the first quarter they might be making only parts below that, i.e. design a new ROM to be capable of 180ns, the first while they're making a lot that only check out at 200, they analyse all the bad ones they can't even sell at 200 to see what's going wrong... they tune up the fab, use sliiightly different doping in the wafer, juggle layer thicknesses, or if they royally screwed up, patch the master, then next quarter they moved the "hump" of the normal curve right over to where they want it right over the 180, or if they're lucky a bit past it. Now most of their production is at 180, but stuff falling in the 200ns grade is only a little slower than 180, not just passing 200. But now they're getting say 20% qualify at 150, woohoo, they can sell those at a premium. Over the span of years their master is wearing out, the hump slides back past 180 again. If they are selling a lot of these and it's really bug free, they may remaster, maybe throw in a trick or two they've learned, have a new master that peaks at 150.

Meanwhile Compaq or someone got on the phone, and said "Hey we like the price and performance is good enough at 200ns, we'll take umpteen thousand 200ns... a month" yay and yeehaw, no flogging the sales reps to push them out, no penny ante single box shipments to minor suppliers, we've landed a big one.... However, it's a bigger order for 200ns than is actually getting marked 200ns now... no problemo, stamp some 180s at 200ns, because their engineers are anal and want that exact qualified part and speed. Could they sell the 180 for higher markup? Yah in theory, but having 180s sitting in inventory waiting for the sales guys to push 'em a few thousand at a time, and the overhead with that doesn't keep the lights on. They are still skimming off the 150 though and things are going good, not quite good enough to market a speed higher than 150, but some of them might test good at 140, a real small handful at 120.

And this crap is going on in waves over all designs over years, so retro youtuber gets in a dozen EPROM from excess compaq stock from a particular time period found NOS on a shelf and says, "This 91xy08923 part is hot shit, marked 200ns and they mostly do 165ns!!!" the original source of a couple of thousand dries up and for years after ppl are buying earlier and later standard production 91xy08923 and wondering why their crap fails because they're getting tight binned 200ns, and should be buying 180ns..

But anyway, trying to characterize the whole production of a particular part number from a single supplier will drive you crazy, but it helps to have some idea of what the speed bins were at the time of production and a little period each way to know whether fab was tuning up, wearing out, or peak productivity.

TL;DR silicon lottery is like a lottery with silicon.

That was an awesome detailed explanation on binning vs market requests etc. Thanks!

ECS MX-381 revision 1.0 based on Macronix MX83C305FC, MX83C306FC

Ever since i tried Octek Jaguar V i have been curious about this chipset and what it is capable of.
Jaguar V was able to scale to 50MHz, but with some compromises, also it has soldered on board processor, which is a limiting factor.
So it was not clear if the chipset works well with SXL2 and BL3 CPUs and where the upper limit is.

A while ago there was Jaguar V revision 2.0 for sale on Ebay, but the price was too high for the level of my curiosity, especially given the fact that other potent looking motherboards based on this chipset show for sale relatively frequently.
Most of them come with clock generators that don't go beyond 80MHz (40MHz FSB), but this one here is with crystal oscillator, which is preferable for overclocking.

The board came in perfect condition.
Notice how there are no scratches on the sides of the ISA slots - a sign that it was plugged somewhere and was used like that for some time.
Then it was retired on a warehouse shelf where it collected that specific thin layer of very fine dust, until it finally arrived here.

This chipset has built-in 8Kb level 2 cache.
Octek's Jaguar V boards have it externalized. Either Octek engineers decided that the built-in one is not good enough, or simply wanted to offer upgrade options for their assemblies, or both.
ECS's board (including the fleet of similar ones) uses the internal L2 buffer, which made me wonder if it will be able to scale past the standard 40MHz.

Hardware components used for the testing:
STB Nitro 2Mb video card (Cirrus Logic GD-5434).
Standard IDE controller with 2Gb industrial Transcend CF card.
16Mb (4x4Mb) of trusted RAM modules with 9 chips each. 3 chip modules make the system unstable.

Hard no lights no matter what past 50MHz FSB (100MHz crystal oscillator).

Tried 3 different BIOSes for this motherboard. Two from AMI and the Mr.BIOS from Octek Jaguar V.
Mr.BIOS is significantly faster than the microcodes from AMI. Used it for testing.

--- IBM BL3 at 100MHz (2x50), ISA bus at 11MHz according to Mr.BIOS

All BIOS settings on max, except:
MEMORY TYPE = 80ns (best is 60ns)
AT-BUS CLOCK = 11MHz (best is 25MHz)
LEVEL 2 CACHE = DISABLED

Level 2 cache must be disabled, or POST never completes. Bummer.
Still, this is faster than 90MHz with level L2 cache enabled.
Apparently the compact 8Kb buffer is a minor factor.

If AT-BUS CLOCK is beyond 12.5MHz, local storage device not recognized.

Other than that things just work.

--- TI SXL2 at 80MHz (2x40), ISA bus at 13.3MHz

Inherently unstable. It is the CPU itself. Need to purchase few more SXL2-66 chips for binning and then retest.

All BIOS settings on max, except:
AT-BUS CLOCK = 13.3MHz (best is 20MHz)

If AT-BUS CLOCK is beyond 13.3MHz, local storage device not recognized.

Level 2 cache is enabled.

Speedsys hangs. Common problem with SXL2 processors.

Sharing numbers only here for the few tests that survived:
Wolf3D: 49.2 fps
Superscape: 32.2 fps
PC Player: 10.7 fps

--- TI SXL2 at 50MHz, ISA bus at 12.5MHz

BIOS settings and side effects exactly like with BL3 at 100MHz above.

Things just work.

--- AMD 386DX at 45MHz, ISA bus at 15MHz

All BIOS settings on max.
Level 2 cache is enabled.

System felt fully stable in DOS and Windows, but surprisingly, the more challenging offline graphics tests failed hard.
Had to step-down to 40MHz for them to complete.
Bummer.

---

Performance-wise the board is slow at interactive graphics, but does pretty ok with FPU heavy tasks - middle of the range.
System is probably affected by the dysfunctional level 2 cache at 50MHz FSB. I can see now why Octek elected to provide alternative solution.
Good chance Octek Jaguar V revision 2.0 (on-board level 2 cache chips and CPU socket) may be the pinnacle of Macronix MX83C305FC, MX83C306FC based motherboards, but that remains to be verified.

benchmark results (look for ECS MX-386)

Gemlight 386DX-3340 based on Symphony 'Haydn II' SL82C362, SL82C461, SL82C465

Several motherboards based on this chipset have been examined already and was determined that they are the fastest clock-to-clock 386 class assemblies.
This one here is no different.

Not far ago i purchased two of these boards for about $50 each. Not a bad price.
The reason i spend the $ and time on them was see if they scale well to 50MHz FSB.
Not all boards based on this chipset can do it, but if these two can - that will be neat.
The type of chips used in them is similar to DTK PEM-0030Y and DTK PEM-0036Y, which can do 50MHz FSB just fine, so that reinforced my decision.

Was not able to find relevant documentation online, but that's fine since the assembly is basically jumperless and the level 2 cache jumper configuration is printed on the PCB.

Supported are IBM BL3, TI 486SXL2 and 386DX processors.
Up to 32Mb FPM RAM.
Default BIOS is rudimentary, but used microcodes from DTK PEM-0036Y - worked fine.
Despite silkscreen prints specify jumper configurations for 64Kb, 128Kb, and 256Kb level 2 cache, only 64Kb is possible. 128/256Kb lead hard to POST not completing or error reported about external cache not recognized.
I don't think there are 8K x 8 cache chips with faster rating than 20ns. I have 3 full sets of such chips. None of them function properly above 40MHz. So, that's the main limiting factor for this motherboard.
50MHz FSB is possible, but without level 2 cache. Obviously performance is not the best.

Decided not to do a full round of testing - it won't reveal anything important.
Basically, at 40MHz everything just works.
Same for 50MHz without level 2 cache.

If overclocking is not an objective, these boards are great.
Despite the smaller level 2 cache buffer they still outdo pretty much everything out there.

And that's all.
Nothing more to say really.

The choice of PCB colour is unappealing and reaks of cheap, thus I'm not surprised 256K and 128K won't work. You can measure how 256K is connected on your working board and attempt to fix the Gemlight.

Did you try 45 MHz FSB?

Even with 64K, it could make for a nice SXL2-80 system. The Haydn II is about 15% faster with the SXL compared to any other 386 I've tested, so the absense of 256K may be acceptable to some people (but not me!).

About wiring-up 256Kb level 2 cache.
I started doing that, but then stopped myself with the question "why spend the time on it ?" and was not able to answer it, so will probably not do it.
I can add the few wires and potentially increase the price of the boards (i don't plan to keep them - spring cleaning coming) but the investment in time and effort for the couple of extra bucks is still not worth it.

About 45MHz.
Yes, i tried. Seems to be working with level 2 cache, but system is fussy. It is the L2 cache chips.

To your last point - yes, if one is not after peak overclocking - these boards are juust great.