The way separate heatsinks from bl3 and p24d/t cpus is by using a knife gradually inserted at one ofvthe edges.
With a bit of pressure and tilting/twisting the knife the heatsink separates without me damaging or scratching the ceramic surface of the cpu.
The more unpleasant part is the cleaning of the hard-rubber-like paste.
It does not dissolve in acetone, so i usually slice it carefully on tin ribbons with new exactoknife and remove them one by one.

Are you doing this with the CPU soldered to a PCB, or already desoldered? I normally use a razorblade and tap that with a hammer, but because the package is QFP, the force will be going on the QFP pins to hold the chip in place. I suppose I could place the heatsink in a vise instead of the PCB, but then the CPU will fall onto the vise, damaging the pins.

Yes you either hold the heatsink or put the opposite corner of the cpu on stable surface.
Thise way there is no force applied whatsoever on the pins holding the cpu to the adapter.

For soldered on board cpus i had to do it only once.
Technique vas similar in general but only insert the knife in ther corner betweem cpu and heatsink then start tiltiing it.
Otherwise if you keep forcing it between cpu/heatsink you can really cause some damage to the pins.

feipoa wrote on 2022-10-16, 21:15:

I'm hesitant to heat up the heatsink again. I don't have enough experience determining the melting point of thermal epoxy.

Probably less than the temperature that would damage a powered down cpu.

It is best to do it when soldered.

An image of solder past application. This is about the most you want to use. I still had one solder bridge I had to desolder, which takes a second with the finest iron tip.

An image of the BL3 soldered on. You can see the epoxy marks to indicate where I normally have a 25 mm fan glued on.

Another image with a 25 mm fan layed on top. I wasn't going to glue it at this point.

The testbed with a SiS Rabbit board. REVTO486 loads. But that's about it at this point. I'm going to swap motherboards.

With the existing 25 mm fan layed atop the 25 mm heatsink, the VRM gets to 46 C. Alternately, if I use a larger 30 mm fan and offset it, the VRM runs at 36 C.

I then put the BL3 hack into a known working system with a BL3, which is my comical CN Tower build. This system has no issues with the BL3 hack. So I will be moving onto a Symphony Haydn2 based motherboard next.

Excellent progress!

Nice. Will you have time to do some voltage/stable speed/temperature testing?

When you said

The testbed with a SiS Rabbit board. REVTO486 loads. But that's about it at this point. I'm going to swap motherboards.

Do you mean you didn't test further or something more ominous than that ?

feipoa wrote on 2022-10-18, 01:43:

With the existing 25 mm fan layed atop the 25 mm heatsink, the VRM gets to 46 C. Alternately, if I use a larger 30 mm fan and o […]
Show full quote

With the existing 25 mm fan layed atop the 25 mm heatsink, the VRM gets to 46 C. Alternately, if I use a larger 30 mm fan and offset it, the VRM runs at 36 C.
BL3_hack_larger_fan_overhang.JPG

I then put the BL3 hack into a known working system with a BL3, which is my comical CN Tower build. This system has no issues with the BL3 hack. So I will be moving onto a Symphony Haydn2 based motherboard next.
BL3_hack_CN_Tower.JPG

feipoa,

I thought that SiS rabbit chipset is limited to 1MB x 9 SIMMs? Yours has 4M x 9 SIMMs in total of 32MB?

And you have a dual tag SRAM cache and cache is 256K?

Thanks and have fun, cheers,

I had intended to update this thread with more information after testing. I tried the BL3 hack on my Symphony board and SiS boards, but after loading the driver, I cannot perform any tasks in DOS. Sometimes, if I type DIR, I will very slowly get a directory listing, but it never returns back to the C:\ prompt. I then put the BL3 hack into my ALi Panda board and it runs fine at 3x75 (boot to Win3.11 and load IE5). I then put the BL3 hack into my spare ALi Panda board, using ISA rather than VLB cards, but I could not get it to run reliably in anything but 1x with L1 disabled. Odd behaviour. Its possible my spare ALi Panda board has some defect, but I was able to run an SXL2-50 in the board just fine. My cased ALi Panada board has a working PS/2 mouse KBC adaption installed, but when I tried such adaption in the spare Panada board, it could not detect the mouse. So I am thinking the spare Panada has issues.

Nonetheless, I felt discouraged by the results. I think some motherboards have some extra circuitry to allow the BL3 hack to work, while others do not. It is not merely a BIOS update. I am using the same BIOS on my Symphony 461 (DTK PEM-4030Y) board that pshipkov is using on this Symphony 461 (DTK PEM-4036Y), yet pshipkov can use his BL3 hack fine. I then ran the Evergreen BL3-75 module in my Symphony 461 board, and it ran just fine. The Evergreen module has a lot of extra circuitry on it. However, clock for clock, the Evergreen BL2-75 was slower than the SXL2-75, like 19 fps vs 20 fps, which leads me to think the Evergreen's circuitry is slowing it down.

My next plan was to put the BL3 hack back into my cased ALi Panda system to see if it would work at 2x33 and 3x33. My Buffalo tower adapter only works up to 3x25. Maybe the absence of 2 dozen sockets will let me run at least 2x33, or preferably 3x33. I doubt it, but I want to try.

Pentiumspeed, my SiS Rabbit has always worked with 32 MB of RAM. The cache is installed according to the manual.

I'm not going to take the CPU off my Blue Lightning board, but I would like to say that you've done a great job with remodelling that PCB to accomodate the voltage regulator. That's excellent that it works with the AM1117 too, does make me wonder why my board has such a powerful regulator instead.
Thank you for the solder paste tip 😀

I know the feeling as well when things don't work quite right - I've been remodelling one of my laptops to mod from DSTN to TFT and when I think things are working great initially, problems have cropped up and it really kills my energy for working on it.

feipoa wrote on 2022-10-19, 22:53:

I had intended to update this thread with more information after testing. I tried the BL3 hack on my Symphony board and SiS boar […]
Show full quote

I had intended to update this thread with more information after testing. I tried the BL3 hack on my Symphony board and SiS boards, but after loading the driver, I cannot perform any tasks in DOS. Sometimes, if I type DIR, I will very slowly get a directory listing, but it never returns back to the C:\ prompt. I then put the BL3 hack into my ALi Panda board and it runs fine at 3x75 (boot to Win3.11 and load IE5). I then put the BL3 hack into my spare ALi Panda board, using ISA rather than VLB cards, but I could not get it to run reliably in anything but 1x with L1 disabled. Odd behaviour. Its possible my spare ALi Panda board has some defect, but I was able to run an SXL2-50 in the board just fine. My cased ALi Panada board has a working PS/2 mouse KBC adaption installed, but when I tried such adaption in the spare Panada board, it could not detect the mouse. So I am thinking the spare Panada has issues.

Nonetheless, I felt discouraged by the results. I think some motherboards have some extra circuitry to allow the BL3 hack to work, while others do not. It is not merely a BIOS update. I am using the same BIOS on my Symphony 461 (DTK PEM-4030Y) board that pshipkov is using on this Symphony 461 (DTK PEM-4036Y), yet pshipkov can use his BL3 hack fine. I then ran the Evergreen BL3-75 module in my Symphony 461 board, and it ran just fine. The Evergreen module has a lot of extra circuitry on it. However, clock for clock, the Evergreen BL2-75 was slower than the SXL2-75, like 19 fps vs 20 fps, which leads me to think the Evergreen's circuitry is slowing it down.

My next plan was to put the BL3 hack back into my cased ALi Panda system to see if it would work at 2x33 and 3x33. My Buffalo tower adapter only works up to 3x25. Maybe the absence of 2 dozen sockets will let me run at least 2x33, or preferably 3x33. I doubt it, but I want to try.

Pentiumspeed, my SiS Rabbit has always worked with 32 MB of RAM. The cache is installed according to the manual.

I guess mine is early motherboard with rabbit chipset using 128K using 8 16k x 8 chips and one tag ram IC but no manual to document second tag socket, there's two more empty sockets, one normal and one smaller. There is setup silkscreened on the board but other than that, that's it. The basic config is 32K, 64K or 128K Need manual to know more.

Cheers,

Thanks for the positive feedback with this hackjob.

I suspect this VRM is best suited to have its Vout tab connected to a larger power plane for heat dissipation. Some VRMs, like the one I sourced for the SXL2 interposer, don't get hot at all for the currents being supplied. It is possible that other BL3 interposer designs sourced their larger VRM's for the same reason.

This is the semi-final implementation. I decided to leave the factory 25 mm heatsink attached for now, and I re-glued a 25 mm fan on top, as shown. I've left the VRM attached with double-sided tape. The trimpot only has one of the leads soldered to a trace, while the other is bodge wire affixed. It also has double-sided tape, but the torque required to turn the screw is too great, so I have to use a finger to hold the trimpot in place while turning the screw.

The final result would have the VRM and trimpot epoxied to the PCB surface, and have the heatsink/fan replaced with 30-32 mm. However, I want to do some further testing on my cased ALi Panda system without the 2-dozen socket stacks first. Then see what I want to do with it. This BL3 hack implementation won't be all that useful if it doesn't work with a broader range of motherboards.

I'm currently running the CPU at the highest voltage possible with this VRM, which is about 3.9 - 4.0 V, depending on how much current is being drawn. The reason was for testing 100+ MHz. If it can only get up to 75 MHz, then I'd take the voltage back down to 3.5 or 3.6 V.

I pulled the socket tower out of my ALi Panda board to test my BL3 hack.

With the BL3 hack interposer, I can run 2x33 MHz reliably. At 3x33, I can at least get into DOS, but DOOM will hang after about 5 seconds running. That is with the 1117 VRM at its max voltage, which varies between 4.15 V and 3.8 V, depending on how much current the CPU is demanding. This happens with many voltage regulators when you are trying to run too high of a voltage for the dropout requirements. Vary currents cause varying voltages. I will order that ultra low drop-out regulator for my BL3 hack when I put in my next digikey order, which will be after I test the custom SXL2 interposers (on route).

At 2x33 MHz, the system requires slower SRAM timings and 3x25 MHz results in a faster system. 2x33 = 4026 realtics, whereas 3x25 = 3790 realtics.

At 2x40 MHz, the system was freaking out, even with settings on slowest.

Next I wanted to see how low of a voltage this mysterious IBM BL2/3 could run reliably at for its rated frequency of 75 MHz. At 3.30 V, I can boot to DOS, but trying to run DOOM resulted in errors. I upped the voltage to 3.45 V and ran DOOM and Windows 3.11 without issue. Because of this, I've left the trimmer set to run the VRM at 3.60 V. Seemed like a safe option given the variances we've seen.

The temperature of the 1117 VRM got up to 62 C at max voltage. It gets to 42 C if I run the BL3 at 3.6 V. I think the fan would need to overhang the heatsink to achieve more suitable cooling.

I do not think the 22 stacked PGA sockets is limiting my Buffalo BL3 adapter to 75 MHz. If I remove the VLB cards and use only ISA, I can remove all the sockets but still not get reliable 66 MHz operation out of it on this very fussy motherboard motherboard. Perhaps the onboard logic is limiting its overclockability on this particular system. I know I've run the Buffalo at 100 MHz in DOS on other systems.

Next steps would be to start testing my box of 386 motherboards to see how far I can take this BL3 hack on a more normal motherboard. I think the onboard logic on the Buffalo BL3 may have its place though, because I was able to run it on my SiS Rabbit at 100 MHz, but couldn't run the BL3 hack whatsoever on the Rabbit, not even at 75 MHz, which we know it is capable of.

Some of my previous results with the Buffalo on the SiS Rabbit in DOOM were:

CPU, ISA freq, gametics
SXL2-55, 11.1, 4669
SXL2-80, 10, 3579
DRx2-66, 11.1, 4657
BL2-66 , 11.1, 3747
BL3-75, 10.0, 4096
BL3-100, 11.1, 3068

From this data, I think it is safe to infer that the BL3 is faster than the SXL clock-for-clock. However, I ran some additional tests on a Symphony 461 motherboard using the Evergreen BL3 adaptor. This board didn't like the BL3 hack at all, so I had to use the Evergreen which has quite a bit of onboard logic. Both CPUs at 75 MHz (2x37.5), ISA at 12.5 MHz, and ET4000AX : SXL2-75 scores 3639 gametics, while the BL2-75 scores 3946 gametics. The BL2/3 should be faster than the SXL2, so what gives? Does this mean the Evergreen's onboard logic is limiting the performance of the IBM Blue Lightning?

Try a external lab psu just for curiosity. Do you have one?

I’ve often wondered how many mhz we are leaving on the table with these small uncomplicated VRMs

Just thinking back to Thermalwrong's photos, and the extra ceramic caps that had been added post-production. Might be worth adding some more capacitance to the 1117 input and output, and some mlcc between some Vcc and Gnd pins. If the CPU has some internal PLL logic then the Vcc supply bouncing around could stop it locking on, maybe.

Nice numbers on the 1117 temperatures. I forgot that even though running Vout at a lower voltage means that there's a bigger voltage drop over the regulator, the current through the CPU also drops, so the total power burnt in the regulator can go down.

I'm not convinced that the issues at 3x33 and 2x40 are necessarily voltage related. Even at 2x33 MHz, the system wouldn't always boot up to completion. At 3x25 MHz, it would always boot up. Thus, I was wondering if the issue is more related to the CPU's ability to multiply FSB's higher than 25 MHz. If you look at the rated voltages for BL3 adaptors, or those Alaris motherboards, aren't they all advertised for either 3x20 or 3x25? Maybe there is a reason we don't see adaptions with 2x33 MHz, and perhaps that reason is the CPU's inability to reliably cope with these higher bus speeds. It could also be my particular ALi Panda board which just doesn't do well with the BL3 with higher bus speeds.

2x40 was much less stable than 3x33. This is surprising because 2x40 MHz is much closer to the rated speed of 3x25 than 3x33. Thus I am also wondering if the CPU's built in PLL handles 3x better than 2x.

I need to pull out more motherboards to test. It may be that 3x30 MHz is a more optimal configuration.

Crosstalk or dirty voltage methinks.

Try the voltage thing.

You want super clean power, test it on a battery. 😀 lithium ion seems to fit the bill.

3.2-4.2v

Plenty of cells in old laptop packs.

Despite my initial feedback in private that BL3 is more clock-to-clock efficient than SXL2, i stand corrected.
Tested some things here in addition to the numbers you posted and yes SXL2 is clock-t0-clock faster than BL3.
What this means - if you cannot get BL3 to climb to 100MHz but run around its normal speeds (that is 75-80 MHz), then SXL2-66 clocked to 80MHz will be the better choice.
That is evident from the numbers you shared and the tests i ran here to double check that.

I looked at my posts and then few more notes here about the previous BL3 experiments.
Initially i had problems with IDE reliability on the DTK PEM-4036Y board.
Had to switch to Promise EIDE Pro ISA controller. For some reason it worked reliably, while 2-3 other controllers like Goldstar, some UMC one, and forgot what was the third one didn't work reliably.
Also, make sure that JP2 on your IDE-to-SD adapter is set in position 1-2 (5V). For some reason if it is set to 2-3 (3.3V) access to the local storage device gets unreliable - hangs during boot - exactly like you are experiencing it DOS prompt shows up but the system is already locked. Or if i manage to get into DOS - write access is very unreliable.

Other than that, from the sound of it the CPU may not have the chops for 80-100 MHz. At least not with the basic air cooling you got.
You mentioned that in the IODATA module you were able to run it by flushing the L1 cache at every tick, but not sure if this is a valid test that proves that CPU can handle OC with L1 cache enabled.
Can you get it to 100MHz with L1 cache disabled btw ? That will be an important clue. For example - the default CPU in Alaris Cougar was just fine at 3x33 with L1 cache disabled, but no chance to go higher at whatever was the default voltage. Once i swapped it with more capable chip - it can do 100MHz with L1 cache enabled, but still not fully reliable unless Peltier cooling is applied.

Another BL3 chip here needs active cooling to get anything above its default 75MHz.

Can cooling be your problem ?
You mentioned something like 46C operational temperature - that is quite a bit for OC this old silicon.

Just thinking out loud.