The selected locations of these decoupling caps are shown in the following image. Also shown are approximate locations used to connect the upper ground ring to the lower ground ring:

For those wanting to replicate this design, a rough BOM would look something like this:

a) a QFP132 to PGA132 interposer board, ideally one which space for the regulator and convenient GND points

b) TPS72501DCQR voltage regulator, SOT223 package

c) 22 uF tantalum capacitor. I'm using 2312 size because that's what I had.

d) 1 uF or 10 uF ceramic capacitor. I'm using 1 uF in 0805, X7R.

e) narrow 1210 caps for the 5V plane. I'm using 2x 100 nF and 2x 10 nF because that's what I had.

f) 11x 100 nF ceramic capacitor, 0603 size, X7R. I'm using 25 or 50 V. 10 are for the QFP leads, the 11th one is for the 5V section near the fan header

g) 124 K-ohm resistor, 0805 size. I'm using RK73H2ATTD1243F.

h) 500 K-ohm Bourns trimmer, 12-turn. 3269X-1-50HLF. Only Mouser had stock last I checked.

i) 25x25x6 mm fan, 5V, 2.2 CFM - if you are using the stock 25 mm heatsink. MF25060V2-1000U-A99

j) epoxy, for fan and trimmer mounting

k) copper tape, for routing the decoupling caps to the QFP leads

l) solder paste. I'm using TS391AX with the finest tip, SMDC-25G

m) 28 AWG, for Vin and Vout only. Also for connecting the 5V plane to itself

n) 30 AWG, for all other connections to the voltage regulator

o) male-to-female machine pin header, for fan

p) Dremel, extra narrow drill bit set, enamel scrapping tool (scalpel), hot air station, soldering iron w/fine tips, etc.

q) heatsink, only if you aren't using the IBM standard 25 mm one. You can probably fit up to a 35 mm heatsink and still have room for the trimmer. I have a 32 mm heatsink I was going to go with if I was using a BL3 CPU without a factory heatsink still mounted.

r) kapton tape, for use as an electrical insulator

s) time, patience, and a steady hand.

First up, remove the 386DX from your interposer using hot air:

Then drill out the 11 holes and vias which connect the PCB's bottom 5V plane from the PCB's top 5V plane. We will be leaving the bottom plane as 5V and making the top plane variable, VCC3.

You will want to reconnect two of the QFP leads as shown below. Start scrapping away the enamel in select locations around the top ground ring, which is used for caps and ground connections. See:

Notice that my first BL3 hack job used interposer Version 2.0, whlie my next hack job is using Version 1.0. In both versions, the top ground ring (actually, a square shape) is not actually connected - it is floating. Was there a reason for doing this? The bottom side of the PCB has the ground ring connected to ground. You will want the top ground ring also connected to ground. Thus, I have scraped away enamel in a dozen or so locations and used 30 AWG to connect the top ground ring.

In drilling the vias, the bottom 5V loop plane gets disconnected from its other half. I reconnected it with 28 AWG in two locations. In the following image, I've also added the narrow 1210 capacitors to the 5 V plane, scrapped some enamel, and have added 3 wires to make more VCC5 connections.

In measuring out the number of GROUND and 5V pins on the PGA-132 package, I count 20x 5V pins and 21x GND pins. However, on these interposers, there are only 8 of 20 5V pins and 10 of 21 ground pins connected. I suppose they did this because the interposer was intended for a 386DX, which presumably draws less current than the IBM BL3. For this reason, I decided to add 3 more 5V wires to the 5V rail, as shown below.

How many VCC5 and GND PGA pins do we need for a BL3 at 100-110 MHz? Assuming we fully utilise the 1 A output of the VRM, and we have 11 VCC5 pins connected, that's 91 mA per pin. If someone has a recommendation for how many PGA VCC5 pins I should connect and why, please let me know.

In connecting the lower ground ring to the upper ring, I am routing 30 AWG thru ground vias. This is to reduce drilling more holes through the PCB. You could selectively drill holes around the ground rings and connect them directly, which might look a little better.

Also shown are the holes drilled for the machine pin fan header and an additional ground hole for the VRM's GND tab:

Here is the view with all ground ring connections being made, minus the GRD wire that goes directly to the VRM's TAB:

Below is the top view of the ground ring connected. Be sure to leave space for soldering capacitors on later. Also shown below are the 28 AWG wire for connecting Vin (5V) to the VRM, and 28 AWG wire for connecting VRM's Vout (VCC3) to the newly made 3V plane. 28 AWG was the largest I had which would fit under the IBM BL3. For what I have on hand, AWG w/cladding measuresments are (diameter):

30 AWG: 0.48 mm - 0.24 mm core - Ampacity at 75C = 0.86 A
28 AWG: 0.56 mm - 0.31 mm core - Ampacity at 75C = 1.4 A
26 AWG: 1.19 mm - 0.38 mm core - Ampacity at 75C = 2.2 A
24 AWG: 1.30 mm - 0.48 mm core - Ampacity at 75C = 3.5 A

You may be able to get away with 26 AWG if you strip off the cladding, but the hassle in insulate it from the other wire was too involved for me. I briefly attempted to look for ampacities for these smaller wires, but I didn't see a chart for ampacity as a function of temperature. The chart here, https://learnmetrics.com/wire-gauge-chart-amp-wire-sizes/ shows ampacity at 75 C. For the BL3 hack, we are up to 47C on the VRM casing, and probably up to 40 C under the CPU. At 75 C, the ampacity of 28 AWG, according to the chart, is 1.4 A. Normally, a wire's ampacity increases with less temperature. Also, you normally try to stay under 80% of a wire's ampacity, which would be 1.12 A. The max VRM output is 1 A, so 28 AWG is probably ideal.

The two long parallel wires shown below are 28 AWG and w/cladding they just fit under the QFP132 CPU and between pins 1 and 132.

Notice how I have scrapped away part of the copper ground for the fan header. This is for the fan's VCC5 connection:

Ensure that you scrap away the enamel for the VRM's GND tab, as well as the centre GND pin for the VRM. I have also surrounded the centre ground pin with kapton tape for added isolation of the VRM's other 4 pins. On my #2 BL3 hack, I used some thermal tape under the VRM body to stick it to the PCB (not shown):

Next, you need to solder it all together in a relatively confined area using 30 AWG. Follow the connection diagram on the datasheet, provided earlier. Luckily, the GND rings make a lot of the connections fairly simple:

For BL3 hack #2, I decided to mount the 1 uF MLCC upright, again utilising the GND ring:

Connect the pins of your trimmer as shown, and set it for about 200 K-ohm.

Next up, solder on the machine pin fan header. You can use square headers if you wish, but I like how Evergreen used these headers on their SXL2 and BL3 upgrade modules. You will also want to mark out your VCC3 locations on the QFP leads for the 100 nF 0603 capacitors. First, lay some kapton tape to cover adjacent vias, then place the copper tape on the CPU's lead as shown. Also ensure you've scrapped away the acrylic enamel on the ground ring for the capacitor, as shown:

Unfortunately, the adhesive on the copper tape isn't all that sticky. So the best means I've found to keep everything in place while soldering on the capacitors, was to use solder paste, as shown:

My method: Place the cap into location, then touch the GND lead of the cap first to melt the paste and use your other hand to keep the capacitor in place using a tooth pick. Just heat it enough so that the capacitor won't move. Same for the other side of the capacitor. This will keep the copper tape held down while you solder the copper tape to the CPU's leads (also use paste there). Once everything isn't floating around any longer, take your iron back to the GND plane with some traditional solder, and solder it down better. Same for the other end of the capacitor.

Another view of capacitor soldering here:

Do the same for the 9 other capacitors:

Be sure to epoxy on the trimmer. The screw on this model of trimmer is rather stiff in comparison to through-hole style trimmers, so be sure to wait for the epoxy to dry fully before testing turning the screw. Also, add contact points for the fan's header on the underside of the PCB. The drill locations are such that VCC5 and GND are in the correct location for easy soldering. Add the 100 nF cap to VCC5 at the fan header's underside, as shown:

The soldering should be complete at this point.

If you aren't running crazy speeds (<80 MHz), you can probably get away without using a fan. I think it looks better without the fan:

Next, I was considering if I need to cool the VRM itself. I was looking into using larger fans over the factory 25 mm heatsink. I tried a 35 mm fan:

And offsetting a 30 mm fan:

But it looked pretty bad. I didn't notice much decrease in surface temperature with the offset 30 mm fan or centred 35 mm fan. They are both fairly low CFM.

At 100 MHz and 3.65 V, I measure a maximum surface temperature on the VRM at 47 C; this was while running DOOM for several minutes.

Shown below are the two BL3 hacks, side-by-side.

On Hack #2, I did a little experimenting to determine if the 10 extra 100 nF caps were necessary. Before soldering them on, I decided to run Hack #2 on my SiS Rabbit motherboard.

BL3 Hack #2
5 KHz noise at 3.8 V, 1x33 MHz:
no caps : 170 mV

But I couldn't get it running faster than 2x40 MHz, even if I upped the voltage to 4.13 V. This same CPU, when it was soldered on the Buffalo interposer, could do 100 MHz at 4.13 V, no problem. Fortunately, after soldering on the 10x 100 nF caps to BL3 Hack #2, 100 MHz worked fine. I'm currently running it at 3.65 V and 100 MHz. So it seems like the extra caps may be necessary.

After adding the 10 caps,

BL3 Hack #2
5 KHz noise at 3.8 V, 10x caps, 1x33 MHz:
105 mV, or 112 mV at 3x33 MHz

5V noise at 3x33 was 66 mV

Notice how engaging clock tripping increase the noise slightly, while on the SXL2 project, clock doubling decreased the noise. It is also curious how the 2-5 KHz waveform is not present with the BL3 CPU, while it always showed itself with the SXL2 CPU.

Next, I compared the noise of Hack #2 to Hack #1,

BL3 Hack #1
5 KHz noise at 3.8 V, 10x caps, 1x33 MHz:
84 mV, or 94 mV at 3x33 MHz

5V noise at 3x33 was 55 mV

The noise measured on Hack #1 is a little better than on Hack #2. I'm not sure if this has to do with different PCB revisions for the interposer.

At 47 C running, I don't think I need a heatsink or fan on the VRM. The datasheet mentions max operating temperature 150 C, or 125 C recommended.

Appreciate it. Thanks !

I ran a few additional tests to support 10x caps vs. 4x caps.

With BL3 hack #1 and 4 caps, 108-112 Mhz would not run at 3.65 V. After adding 6 more caps, for a total of 10 caps, it can run at 112.5 MHz at 3.65 V. Similarly, with 4 caps, 2x50 needed 4.1 V, now it can do with only 4.0 V.

At 120 Mhz, I previously could not get DOOM to complete w/4 caps. With 10 caps now, I can get DOOM to complete at 4.2 V. it scores 26.0 fps w/ISA at 13.3 Mhz. 112.5 MHz gets 24.4 fps and 90 Mhz gets 23.9 fps.

With BL3 hack #2, I only ran a few values, namely 3x33 and 3x37.5. 3x33 has no issue at 3.5 V (I did not test lower) but I had to run 3.9 V for 112 MHz to be stable. Hack #1 can do 112 Mhz at 3.65 V (I did not try lower). I'm not sure if Hack #2 requires a higher voltage because it showed a little more VCC3 noise compared to #1.

I ran the regulator up to 4.78 V. I then trimmed it higher, but the system hang up. I think the regulator goes into some kind of bypass or shutdown mode. I did measure 4.92 V on Vout though.

Very nice.
It looks to me that the improvements can boost less potent chips that struggle to reach 90/100MHz.
Will read/comprehend properly in the coming days. Expect questions.

Yeah, it is very exciting news. On Hack #2, the CPU could barely do 80 MHz at 4.1 V, now it can do 112 MHz at 3.6 V. I'm surprised you did not witness a similar outcome was not observed on the SXL2 interposer. This was not a condition I tested for on the SXL.

<duplicate post>
<deleted>

Intriguing... I notice on my SL4100 BL3 cobalt lpx board, IBM stuck the VRM right next to the CPU to keep down the transients. I am not seeing it looking super "cappy" around the CPU though, so there may be potential for improvement.

I feel like I'd kinda be in caveman mode on that 5khz though. "Who make kilohertz noise? Fan make kilohertz noise? Fan get big cap. Now who make kilohertz noise? HDD motor make kilohertz noise?... "

No. The sources of noise comes from current spikes transits as transistors switches on and off all over the silicon die of any CPU and chipsets, Looks like white noise on scope (wide band of high frequency spikes) same like you would hear from old fashioned TV with snow on a non signal channel.

Sine wave noise overlaid the white noise is from somewhere else, not the CPU or chipset. Usually improper ground clip connected with the scrope or power supply passing low freq sine wave into supply rails.

This what you should expect to see for a regular white noise: Lots of high current ripple rated capacitors (ultra low ESR caps and more MLCC ceramic capacitors) squeezes the band more thinner on vcore rails for CPU for example helps with overclocking too. These MLCC ceramic must be located very close to the CPU socket area by few mm and always preferably right at each Vcore pin with strong ground return, same thing with any every chipsets voltage rails right at each pin 1 to 4mm apart to a MLCC capacitor. Datasheets for CPU and chipset specifies minimum ripple voltage that includes the low frequency ripple too.

https://www.google.ca/search?q=white+noise+on … =jLKyf_aHPHv1MM

Ripple overlaid this normal noise is a real concern, this is low frequency ripple, causes overall stability or scope probe connected wrong or linear regulator oscillating due to not following the datasheet requirements. This is corrected with rebuilding, redesign the linear regulator circuit or replace power supply and more good ESR electrolytic capacitors for bulk filtering to flatten the ripple, NOT the ceramic capacitors. https://images.app.goo.gl/wrvrUCJTdfq713MC6

Sources of thinking noise from sound card is due to this especially if the motherboard and cards PCB is 2 layer and poorly designed. 4 or more layers is preferred and well designed helps here for stability like overclocking and cut down noise getting into audio.

Back in the day in high school when learning to build TTL circuit I caught TTL oscillating at high frequency when signal voltage was in the switching voltage range.

Cheers,

BitWrangler wrote on 2023-03-20, 19:07:

I feel like I'd kinda be in caveman mode on that 5khz though. "Who make kilohertz noise? Fan make kilohertz noise? Fan get big cap. Now who make kilohertz noise? HDD motor make kilohertz noise?... "

I'm not sure which setup you are referring to concerning this thread. I noted a few posts ago that I did not get any low frequency noise with this BL3 CPU hack job. See:

feipoa wrote on 2023-03-19, 13:08:

It is also curious how the 2-5 KHz waveform is not present with the BL3 CPU, while it always showed itself with the SXL2 CPU.

I only witnessed this noise using the custom SXL2 interposer. It didn't matter which regulator, caps, motherboard, fan, no fan, etc. I used, it was always present, however did not appear to impact performance. The Vp-p was low. But this is a topic for the other thread.